Weekly TMLR digest for Jun 28, 2026
TMLR
New certifications
==================
Survey Certification: Hardware Acceleration for Neural Networks: A Comprehensive Survey
BIN XU, Ayan Banerjee, Sandeep Gupta
https://openreview.net/forum?id=Da8LO5NvDU
---
Survey Certification: A Survey of Flow Matching in Reinforcement Learning
Nabuat Zaman Nahim, Fairoz Nower Khan, Peizhong Ju
https://openreview.net/forum?id=P6e5IC4gPe
---
J2C Certification: DAG-aware GAT for Causal Effect Estimation
Manqing Liu, David Remy Bellamy, Andrew Beam
https://openreview.net/forum?id=pzFwkjRVXy
---
J2C Certification: Beyond the Linear Separability Ceiling: Aligning Representations in VLMs
Enrico Vompa, Tanel Tammet, Mohit Vaishnav
https://openreview.net/forum?id=3uX4p80bN0
---
Accepted papers
===============
Title: VerlTool: Towards Holistic Agentic Reinforcement Learning with Tool Use
Authors: Dongfu Jiang, Yi Lu, Zhuofeng Li, Zhiheng Lyu, Ping Nie, Haozhe Wang, Alex Su, Hui Chen, Kai Zou, Chao Du, Tianyu Pang, Wenhu Chen
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated success in enhancing LLM reasoning capabilities, but remains limited to single-turn interactions without tool integration. While recent **A**gentic **R**einforcement **L**earning with **T**ool use (ARLT) approaches have emerged to address multi-turn tool interactions, existing works develop task-specific codebases that suffer from fragmentation, synchronous execution bottlenecks, and limited extensibility across domains. These inefficiencies hinder broader community adoption and algorithmic innovation. We introduce **VerlTool**, a unified and modular framework that addresses these limitations through systematic design principles. VerlTool provides four key contributions: **(1)** upstream alignment with VeRL ensuring compatibility and simplified maintenance, **(2)** unified tool management via standardized APIs supporting diverse modalities including code execution, search, SQL databases, and vision processing, **(3)** asynchronous rollout execution achieving near 2
speedup by eliminating synchronization bottlenecks, and **(4)** a comprehensive evaluation demonstrating competitive performance across 6 ARLT domains. Our framework formalizes ARLT as multi-turn trajectories with multi-modal observation tokens (text/image/video), extending beyond single-turn RLVR paradigms. We train and evaluate models on mathematical reasoning, knowledge QA, SQL generation, visual reasoning, web search, and software engineering tasks, achieving results comparable to specialized systems while providing a unified training infrastructure. The modular plugin architecture enables rapid tool integration requiring only lightweight Python definitions, significantly reducing development overhead and providing a scalable foundation for tool-augmented RL research.
URL: https://openreview.net/forum?id=g2LCOW43Md
---
Title: Deep Sprite-based Image Models: An Analysis
Authors: Zeynep Sonat Baltaci, Romain Loiseau, Mathieu Aubry
Abstract: While foundation models drive steady progress in image segmentation and diffusion algorithms compose always more realistic images, the seemingly simple problem of identifying recurrent patterns in a collection of images remains very much open. In this paper, we focus on sprite-based image decomposition models, which have shown some promise for clustering and image decomposition and are appealing because of their high interpretability. These models come in different flavors, need to be tailored to specific datasets, and struggle to scale to images with many objects. We dive into the details of their design, identify their core components, and perform an extensive analysis on clustering benchmarks. We leverage this analysis to propose a deep sprite-based image decomposition method that performs on par with state-of-the-art unsupervised class-aware image segmentation methods on the standard CLEVR benchmark, scales linearly with the number of objects, identifies explicitly object categories, and fully models images in an easily interpretable way.
URL: https://openreview.net/forum?id=pXuxMLFo9g
---
Title: Hardware Acceleration for Neural Networks: A Comprehensive Survey
Authors: BIN XU, Ayan Banerjee, Sandeep Gupta
Abstract: Neural networks have become a dominant computational workload across cloud and edge platforms, but their rapid growth in model size and deployment diversity has exposed hardware bottlenecks that are increasingly dominated by memory movement, communication, and irregular operators rather than peak arithmetic throughput. This survey reviews the current technology landscape for hardware acceleration of deep learning, spanning Graphics Processing Units (GPUs) and tensor-core architectures, domain-specific accelerators (e.g., Tensor Processing Units (TPUs)/Neural Processing Units (NPUs)), Field-Programmable Gate Array (FPGA)-based designs, Application-Specific Integrated Circuit (ASIC) inference engines, and emerging Large Language Model (LLM)-serving accelerators such as Language Processing Units (LPUs), alongside in-/near-memory computing and neuromorphic/analog approaches. We organize the survey using a unified taxonomy across (i) workloads (Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Graph Neural Networks (GNNs), Transformers/Large Language Models (LLMs)), (ii) execution settings (training vs.\ inference; datacenter vs.\ edge), and (iii) optimization levers (reduced precision, sparsity and pruning, operator fusion, compilation and scheduling, and memory-system/interconnect design). We synthesize key architectural ideas such as systolic arrays, vector and Single Instruction, Multiple Data (SIMD) engines, specialized attention and softmax kernels, quantization-aware datapaths, and high-bandwidth memory, and we discuss how software stacks and compilers bridge model semantics to hardware. Finally, we highlight open challenges—including efficient long-context LLM inference (Key-Value (KV)-cache management), robust support for dynamic and sparse workloads, energy- and security-aware deployment, and fair benchmarking—pointing to promising directions for the next generation of neural acceleration.
URL: https://openreview.net/forum?id=Da8LO5NvDU
---
Title: LMA: Latent Motion Adjuster for Physics-based Multi-agent Interaction
Authors: Taiki Yano, Satoshi Yagi, Satoshi Yamamori, Jun Morimoto
Abstract: Learning interactive multi-agent behaviors from scratch is often sample-inefficient and fails to exploit reusable skills learned in simpler settings. While latent skill representations enable efficient single-agent reinforcement learning, their extension to multi-agent interaction requires conditioning behaviors on other agents without destroying pretrained structure. We formulate multi-agent interaction as a latent adaptation problem and propose the Latent Motion Adjuster (LMA), a lightweight conditional module that modifies latent actions produced by a pretrained single-agent policy based on other agents’ states. Rather than relearning policies from scratch, our method performs structured residual adaptation in latent space, enabling efficient skill reuse under both cooperative and competitive scenarios. Experiments on physics-based control benchmarks demonstrate that latent-space adaptation improves sample efficiency and interaction performance over fine-tuning and strategic baselines. These results suggest that conditional latent modulation provides a principled mechanism for transferring single-agent skills to multi-agent reinforcement learning.
URL: https://openreview.net/forum?id=OR5ClOTaWU
---
Title: Continual Memorization of Factoids in Language Models
Authors: Howard Chen, Jiayi Geng, Adithya Bhaskar, Dan Friedman, Danqi Chen
Abstract: As new knowledge rapidly accumulates, language models (LMs) with pretrained knowledge quickly become obsolete. A common approach to updating LMs is fine-tuning them directly on new knowledge. However, recent studies have shown that fine-tuning for memorization may be ineffective in storing knowledge or may even exacerbate hallucination—raising doubts about its reliability when applied repeatedly. To study this, we formalize the problem of continual memorization, where a model must memorize and retain a set of factoids through multiple stages of fine-tuning on subsequent datasets. We first characterize the forgetting patterns through extensive experiments and show that LMs widely suffer from forgetting, especially when needing to memorize factoids in the second stage. We posit that forgetting stems from suboptimal training dynamics which fails to: (1) protect the memorization process when learning factoids or (2) reduce interference from subsequent training stages. To test this hypothesis, we explore various data mixing strategies to alter the fine-tuning dynamics. Intriguingly, we find that mixing randomly generated word sequences or generic data sampled from pretraining corpora at different training stages effectively mitigates forgetting (REMIX: Random and Generic Data Mixing). REMIX can recover performance from severe forgetting, outperforming replay methods and other continual learning baselines. We analyze how data mixing can influence the learning process and find that robust memorization follows a distinct pattern—the model stores factoids in earlier layers than usual and diversifies the layers that retain them, which results in easier recall and manipulation of the learned factoids.
URL: https://openreview.net/forum?id=5Yd5QAKIFR
---
Title: SpurLens: Automatic Detection of Spurious Cues in Multimodal LLMs
Authors: Parsa Hosseini, Sumit Nawathe, Mazda Moayeri, Sriram Balasubramanian, Soheil Feizi
Abstract: Unimodal vision models are known to rely on spurious correlations, but it remains unclear to what extent Multimodal Large Language Models (MLLMs) exhibit similar biases despite language supervision. In this paper, we investigate spurious bias in MLLMs and introduce SpurLens, a pipeline that leverages GPT-4 and open-set object detectors to automatically identify spurious visual cues without human supervision. Our findings reveal that spurious correlations cause two major failure modes in MLLMs: (1) over-reliance on spurious cues for object recognition, where removing these cues reduces accuracy, and (2) object hallucination, where spurious cues amplify the hallucination by over 10x.
We investigate various MLLMs and datasets, and validate our findings with multiple robustness checks. Beyond diagnosing these failures, we explore potential mitigation strategies, such as prompt ensembling and reasoning-based prompting, and conduct ablation studies to examine the root causes of spurious bias in MLLMs. By exposing the persistence of spurious correlations, our study calls for more rigorous evaluation methods and mitigation strategies to enhance the reliability of MLLMs.Unimodal vision models are known to rely on spurious correlations, but it remains unclear to what extent Multimodal Large Language Models (MLLMs) exhibit similar biases despite language supervision. In this paper, we investigate spurious bias in MLLMs and introduce SpurLens, a pipeline that leverages GPT-4 and open-set object detectors to automatically identify spurious visual cues without human supervision. Our findings reveal that spurious correlations cause two major failure modes in MLLMs: (1) over-reliance on spurious cues for object recognition, where removing these cues reduces accuracy, and (2) object hallucination, where spurious cues amplify the hallucination by over 10x.
We investigate various MLLMs and datasets, and validate our findings with multiple robustness checks. Beyond diagnosing these failures, we explore potential mitigation strategies, such as prompt ensembling and reasoning-based prompting, and conduct ablation studies to examine the root causes of spurious bias in MLLMs. By exposing the persistence of spurious correlations, our study calls for more rigorous evaluation methods and mitigation strategies to enhance the reliability of MLLMs.
URL: https://openreview.net/forum?id=jxw6t5VVNL
---
Title: Self-Improvement as Coherence Optimization: A Theoretical Account
Authors: Tianyi Alex Qiu, Ahmed Ismail, Zhonghao He, Shi Feng
Abstract: Can language models improve their accuracy without external supervision? Methods such as debate, bootstrap, and internal coherence maximization achieve this surprising feat, even matching golden finetuning performance. Yet why they work remains theoretically unclear. We show that they can all be understood as *coherence optimization*, the search for a context-to-behavior mapping that is most compressible and jointly predictable, with debate an exact instance and bootstrap and internal coherence maximization closely related to it. We prove that coherence optimization is equivalent to description-length regularization, and that among all such regularization schemes, coherence regularization with a prior derived from a pretrained model optimizes a lower bound of worst-case accuracy for semi-supervised learning (Theorem 5.5). Our theory, supported by preliminary experiments, explains why feedback-free self-improvement works and predicts when it should succeed or fail.
URL: https://openreview.net/forum?id=nR47qAX9oL
---
Title: Behavior-dLDS: A decomposed linear dynamical systems model for neural activity partially constrained by behavior
Authors: Eva Yezerets, En Yang, Misha B. Ahrens, Adam Shabti Charles
Abstract: Brain-wide recordings of large-scale networks of neurons now provide an unprecedented view into how the brain drives behavior. However, brain activity contains both information directly related to behavior as well as the potential for many internal computations. Moreover, observable behavior is executed not only by the brain, but also by the spinal cord and peripheral nervous system. Behavior is a coarse-grained product of neural activity, and we thus take the view that it can be best represented by lower-dimensional latent neural dynamics. Capturing this indirect relationship while disambiguating behavior-generating networks from internal computations running in parallel requires new modeling approaches that can embody the parallel and distributed nature of large-scale neural populations. We thus present behavior-decomposed linear dynamical systems (b-dLDS) to disentangle simultaneously recorded subsystems and identify how the latent neural subsystems relate to behavior. We demonstrate the ability of b-dLDS to decouple behavioral vs. internal computations on controlled, simulated data, showing improvements over a state-of-the-art model that uses behavior to supervise all dynamics based on behavior. We also demonstrate b-dLDS's interpretability benefits on a task-driven RNN dataset featuring a nonlinear relationship between behavior and activations. We then show that b-dLDS can further scale up to tens of thousands of neurons by applying our model to a large-scale recording of a zebrafish hindbrain during the complex positional homeostasis behavior, wherein b-dLDS highlights asymmetry in behavior-related dynamic connectivity networks.
URL: https://openreview.net/forum?id=8p9tP9qLPN
---
Title: Code Reasoning for Software Engineering Tasks: A Survey and A Call to Action
Authors: Saurabh Pujar, Ira Ceka, Irene L Manotas, Gail Kaiser, Baishakhi Ray, Shyam Ramji
Abstract: The rise of large language models (LLMs) has led to dramatic improvements across a wide range of natural language tasks. Their performance on certain tasks can be further enhanced by incorporating test-time reasoning techniques. These inference-time advances have been adopted into the code domain, enabling complex software engineering (SWE) tasks such as code generation, test generation and issue resolution. However, the impact of different reasoning techniques on code-centric SWE tasks has not been systematically explored. In this work, we survey code reasoning techniques that underpin these capabilities, with a focus on test-time compute and inference-time reasoning paradigms. We examine a variety of code-specific reasoning methods and progressively build up to SWE agents, which combine planning, tool use, and multi-step interaction. We also compare the impact of different techniques on coding tasks, highlighting their relative importance and outlining open challenges and future research directions. Across commonly used models and benchmarks, we find that approaches exploiting code-specific signals (e.g., structure and execution feedback) are frequently associated with improved performance, motivating a dedicated study of code reasoning beyond natural-language reasoning. Our contributions are: (1) to the best of our knowledge, the first dedicated survey of code reasoning for SWE tasks, highlighting overarching reasoning strategies, hybrid methods, and agentic approaches; (2) a taxonomy of inference-time techniques used to drive code reasoning, accompanied by a curated set of under-explored benchmarks with high potential for SWE evaluation; (3) a comparative analysis of reasoning design patterns across commonly used models and benchmarks; and (4) a synthesis of gaps in current methods and evaluation practices, identifying under-explored areas and concrete opportunities for future research.
URL: https://openreview.net/forum?id=zZa3u6LKwO
---
Title: Tabular Learning Revisited: An Empirical Study of Tabular Classification
Authors: Guri Zabërgja, Arlind Kadra, Christian Frey, Josif Grabocka
Abstract: Tabular data represent one of the most prevalent data formats in applied machine learning, largely because they accommodate a broad spectrum of real-world problems.
Existing literature has studied many of the shortcomings of neural architectures on tabular data and has repeatedly confirmed the scalability and robustness of gradient-boosted decision trees across varied datasets. However, recent deep learning models have not been subjected to a comprehensive evaluation under conditions that allow for a fair comparison with existing classical approaches. This situation motivates an investigation into whether recent deep-learning paradigms outperform classical ML methods on tabular data. Our survey fills this gap by benchmarking twenty state-of-the-art methods, spanning neural networks, classical ML and AutoML techniques. Our empirical results over 68 diverse classification datasets from a well-established benchmark indicate a paradigm shift, where Deep Learning methods outperform classical approaches.
URL: https://openreview.net/forum?id=I8BIGp4XOb
---
Title: Time-Aware Prior Fitted Networks for Zero-Shot Forecasting with Exogenous Variables
Authors: Andres Potapczynski, Ravi Kiran Selvam, Tatiana Konstantinova, Malcolm Wolff, Kin G. Olivares, Ruijun Ma, Michael W. Mahoney, Andrew Gordon Wilson, Boris N. Oreshkin, Dmitry Efimov
Abstract: In many time series forecasting settings, the target series is accompanied by exogenous covariates such as promotions and prices in retail demand, temperature in energy load, calendar and holiday indicators in traffic or sales, and grid load or fuel costs in electricity pricing. Ignoring these exogenous signals can substantially degrade forecasting accuracy, particularly when they drive spikes, discontinuities, or regime and phase changes in the target series. Most current time series foundation models (e.g., Chronos, Sundial, TimesFM, TimeMoE, TimeLLM, and LagLlama) ignore exogenous covariates and make forecasts solely from the numerical history of the target series, thereby limiting their performance. In this paper, we develop ApolloPFN, a prior-data fitted network (PFN) that is time-aware (unlike prior PFNs) and natively incorporates exogenous covariates (unlike prior univariate forecasters). Our design introduces two major advances: (i) a synthetic data generation framework that injects realistic temporal patterns, structural changes, and exogenous dependencies into the PFN prior; and (ii) time-aware architectural modifications that embed inductive biases needed to exploit temporal context. We demonstrate that ApolloPFN outperforms existing baselines across several forecasting benchmarks with exogenous information, including M5, electricity price forecasting, UCI Air Quality, and Solar Energy datasets.
URL: https://openreview.net/forum?id=nJARpxp3cF
---
Title: Reference-Guided Identity Preserving Face Restoration
Authors: Mo Zhou, Keren Ye, Viraj Shah, Kangfu Mei, Mauricio Delbracio, Peyman Milanfar, Vishal M. Patel, Hossein Talebi
Abstract: Preserving face identity is a critical yet persistent challenge in diffusion-based image restoration. While reference faces offer a path forward, existing methods typically suffer from partial reference information and inefficient identity losses. This paper introduces a novel approach that directly solves both issues, involving three key contributions: 1) Composite Context, a representation that fuses high- and low-level facial information to provide ensembled guidance than traditional singular representations, 2) Hard Example Identity Loss, a novel loss function that uses the reference face to address the identity learning inefficiencies of the standard identity loss, 3) Training-free multi-reference inference, a new method that leverages multiple references for restoration, despite being trained with only a single reference. The proposed method demonstrably restores high-quality faces and achieves state-of-the-art identity preserving restoration on benchmarks such as FFHQ-Ref and CelebA-Ref-Test, consistently outperforming previous work.
URL: https://openreview.net/forum?id=g9YzUDnUUS
---
Title: Robust Cross-Domain Alignment
Authors: Anish Chakrabarty, ARKAPRABHA BASU, Swagatam Das
Abstract: The Gromov-Wasserstein (GW) distance is an effective measure of alignment between distributions supported on distinct ambient spaces. Calculating essentially the mutual departure from isometry, it has found vast usage in domain translation and network analysis. It has long been shown to be vulnerable to contamination in the underlying measures. All efforts to introduce robustness in GW have been inspired by similar optimal transport (OT) techniques, which predominantly advocate partial mass transport or unbalancing. In contrast, the cross-domain alignment problem, being fundamentally different from OT, demands specific solutions to tackle diverse applications and contamination regimes. Deriving from robust statistics, we discuss three contextually novel techniques to robustify GW and its variants. For each method, we explore metric properties and robustness guarantees along with their co-dependencies and individual relations with the GW distance. For a comprehensive view, we empirically validate their superior resilience to contamination under real machine learning tasks against state-of-the-art methods.
URL: https://openreview.net/forum?id=0mchjaZZi4
---
Title: LLM-Guider: A Language-Guided Discovery of Symbolic Pruning Metrics for Post-Training Sparsity in LLMs
Authors: Jędrzej Hasiura, Prashant Shivaram Bhat, Elahe Arani, Bahram Zonooz
Abstract: Large Language Models (LLMs) have achieved remarkable advancements in natural language understanding, yet their mammoth size, coupled with substantial training and inference costs, can make them difficult to use in environments with limited resources. To address both memory and efficiency concerns, post-training unstructured sparsity techniques have emerged, focusing on developing optimal pruning criteria to eliminate redundant weights while maintaining performance. However, these approaches often rely on manually crafted pruning criteria, leading to sub-optimal solutions due to heuristic oversimplifications. Therefore, we introduce LLM-Guider, a language-guided symbolic formula optimization framework that seeks to discover optimal pruning criteria through a transparent and systematic process. LLM-Guider comprises three interrelated stages: example selection, formula generation, and formula evaluation, which collectively enable the efficient exploration of the formula space. In addition, LLM-Guider enables the incorporation of intuition, domain, and mathematical knowledge through role prompts, hints, and in-context examples. We also extend the standard set of aggregation strategies over a calibration dataset, resulting in never-seen-before pruning metrics.
Through extensive experiments, we demonstrate that formulas discovered through LLM-Guider is able to find formulas that outperform established baselines.
URL: https://openreview.net/forum?id=SlVQxEiYnY
---
Title: CADO: From Imitation to Cost Minimization for Heatmap-based Solvers in Combinatorial Optimization
Authors: Hyungseok Song, Deunsol Yoon, Kanghoon Lee, Han-Seul Jeong, Soonyoung Lee, Woohyung Lim
Abstract: Heatmap-based solvers have emerged as a promising paradigm for Combinatorial Optimization (CO). However, we argue that the dominant Supervised Learning (SL) training paradigm suffers from a systematic objective mismatch: minimizing imitation loss (e.g., cross-entropy) does not guarantee solution cost minimization. We dissect this mismatch into two deficiencies: Decoder-Blindness (being oblivious to the non-differentiable decoding process) and Cost-Blindness (prioritizing structural imitation over solution quality). We empirically characterize these intrinsic flaws and show that they impose a performance limitation. To overcome this limitation, we propose CADO (Cost-Aware Diffusion models for Optimization), a streamlined Reinforcement Learning fine-tuning framework that formulates the diffusion denoising process as an MDP to directly optimize the post-decoded solution cost. We introduce Label-Centered Reward, which repurposes ground-truth labels as unbiased baselines rather than imitation targets, and Hybrid Fine-Tuning for parameter-efficient adaptation. CADO achieves state-of-the-art performance across diverse benchmarks, showing that objective alignment significantly improves the quality of SL-trained heatmap solvers.
URL: https://openreview.net/forum?id=fvxx5FOED6
---
Title: Numerical Analysis of HiPPO-LegS ODE for Deep State Space Models
Authors: Jaesung R. Park, Jaewook J. Suh, Youngjoon Hong, Ernest K. Ryu
Abstract: In deep learning, the recently introduced state space models utilize HiPPO (High-order Polynomial Projection Operators) memory units to approximate continuous-time trajectories of input functions using ordinary differential equations (ODEs), and these techniques have shown empirical success in capturing long-range dependencies in long input sequences. However, the mathematical foundations of these ODEs, particularly the singular HiPPO-LegS (Legendre Scaled) ODE, and their corresponding numerical discretizations remain unsettled. In this work, we fill this gap by establishing that HiPPO-LegS ODE is well-posed despite its singularity, albeit without the freedom of arbitrary initial conditions. Further, we establish convergence of the associated numerical discretization schemes for Riemann integrable input functions.
URL: https://openreview.net/forum?id=83dhVASBPn
---
Title: A Survey of Flow Matching in Reinforcement Learning
Authors: Nabuat Zaman Nahim, Fairoz Nower Khan, Peizhong Ju
Abstract: Flow Matching (FM) has recently emerged as a principled and efficient generative modeling framework for reinforcement learning (RL), enabling expressive, multimodal policy parameterizations via deterministic probability transport.
Compared to diffusion-based policies that rely on stochastic denoising chains, FM uses sampling based on ordinary differential equations (ODEs), with learned velocity fields, which can substantially reduce inference latency and simplify the incorporation of RL objectives.
As research in flow-based RL rapidly accelerates across offline continuous control, online fine-tuning, and foundation model alignment, the literature has become highly fragmented. In this survey, we provide a comprehensive taxonomy of flow-matching approaches in reinforcement learning. We organize the literature along two axes: the target distribution being modeled (e.g., action policies, value critics, transition dynamics) and the mechanism of RL signal integration (e.g., energy-weighted regression, flow-based policy gradients, and group relative policy optimization).
Furthermore, we survey emerging frontiers such as discrete and non-Euclidean action spaces, provide a systematic comparative analysis against Gaussian and diffusion baselines, and outline critical open problems. Ultimately, this survey serves as a foundational roadmap for the next generation of generative reinforcement learning and alignment.
URL: https://openreview.net/forum?id=P6e5IC4gPe
---
Title: Bridging Mechanistic Interpretability and Prompt Engineering with Gradient Ascent for Interpretable Persona Control
Authors: Harshvardhan Saini, Yiming Tang, Dianbo Liu
Abstract: Controlling emergent behavioral personas (e.g., sycophancy, hallucination) in Large Language Models (LLMs) is critical for AI safety, yet remains a persistent challenge. Existing solutions face a dilemma: manual prompt engineering is intuitive but unscalable and imprecise while automatic optimization methods are effective but operate as "black boxes" with no interpretable connection to model internals. We propose a novel framework that adapts gradient ascent to LLMs, enabling targeted prompt discovery. In specific, we propose two methods, RESGA and SAEGA, that both optimize randomly initialized prompts to achieve better aligned representation with an identified persona direction. We introduce fluent gradient ascent to control the fluency of discovered persona steering prompts. We demonstrate RESGA and SAEGA’s effectiveness across Llama 3.1, Qwen 2.5, and Gemma 3 for steering three different personas, sycophancy, hallucination, and myopic reward. Crucially, on sycophancy, our automatically discovered prompts achieve significant improvement (49.90% compared with 79.24%). By grounding prompt discovery in mechanistically meaningful features, our method offers a new paradigm for controllable and interpretable behavior modification. We release our scripts for RESGA and SAEGA in this github repo: https://github.com/HarshSaini10/RESGA_SAEGA.
URL: https://openreview.net/forum?id=dcmHPxgo4c
---
Title: RobustMAD: Evaluating Real-World Robustness of Multimodal Small Language Models for Deployable Anomaly Detection Assistants
Authors: Anushiya Arunan, Xin Li, Yan Qin, U-Xuan Tan, VUONG NHU KHUE, Xiaoli Li, Chau Yuen
Abstract: Multimodal industrial anomaly inspection assistants are a critical component of next-generation smart factories, enabling interactive vision–language–based querying. However, multimodal large language models remain impractical for on-site deployment due to prohibitive computational demands and privacy risks from cloud-based inference. Compact multimodal \textit{small} language models (MSLMs) offer a deployable alternative, yet progress is constrained by the lack of comprehensive robustness analyses and meaningfully challenging benchmarks that reflect real-world industrial conditions. To address this gap, we develop RobustMAD, the first deployment-motivated benchmark, designed to comprehensively evaluate model robustness through diverse open-ended queries spanning object understanding, anomaly detection, unanswerable problems, and visual quality degradations. Contrary to conventional assumptions, top-performing MSLMs exhibit promising capabilities, surprisingly outperforming even the larger GPT-5 Nano. However, they still fall short of safety-critical requirements, and RobustMAD reveals critical robustness gaps that pose operational risks. In particular, three recurring failure modes emerge: (i) fragile multimodal grounding under fine-grained distinctions or degraded visual conditions, (ii) insufficiently comprehensive responses, and (iii) weak logical grounding on unanswerable or ill-posed queries, leading to hallucinated outputs. Grounded in these insights, we provide actionable guidance for the design of next-generation multimodal industrial inspection assistants that leverage their promising competence.
URL: https://openreview.net/forum?id=skrA9UYNIZ
---
Title: NoisyCoconut: Counterfactual Consensus via Latent Space Reasoning
Authors: Michael M. Jerge, David Evans
Abstract: This paper presents NoisyCoconut, a novel inference-time method that enhances large language model (LLM) reliability using multiple executions with manipulated internal representations. Unlike fine-tuning methods that require extensive retraining, NoisyCoconut operates directly on model representations during inference and requires no retraining. Rather than training models to reason in latent space, we inject controlled noise into latent trajectories to generate diverse reasoning paths. Agreement among these paths provides a confidence signal, enabling models to abstain when uncertain. We demonstrate that this approach achieves effective coverage-accuracy tradeoffs across multiple reasoning benchmarks, providing a practical pathway to improving the reliability of LLM outputs without requiring any retraining. Our experiments show that unanimous agreement among noise-perturbed paths reduces error rates from 40--70% to below 15%, enabling models to exceed 95\% accuracy on mathematical reasoning tasks through selective abstention. Our code is available at https://github.com/mmjerge/noisycoconut.
URL: https://openreview.net/forum?id=5aatZPiCv8
---
Title: Maximizing Confidence Alone Improves Reasoning
Authors: Mihir Prabhudesai, Lili Chen, Alex Ippoliti, Katerina Fragkiadaki, Hao Liu, Deepak Pathak
Abstract: Reinforcement learning (RL) has enabled machine learning models to achieve significant advances in many fields. Most recently, RL has empowered frontier language models to solve challenging math, science, and coding problems. However, central to any RL algorithm is the reward function, and reward engineering is a notoriously difficult problem in any domain. In this paper, we propose RENT: Reinforcement Learning via Entropy Minimization -- a fully unsupervised RL method that requires no external reward or ground-truth answers, and instead uses the model's entropy of its underlying distribution as an intrinsic reward. We find that by reinforcing the chains of thought that yield high model confidence on its generated answers, the model improves its reasoning ability. In our experiments, we showcase these improvements on an extensive suite of commonly-used reasoning benchmarks, including GSM8K, MATH500, AMC, AIME, and GPQA, and models of varying sizes from the Qwen, Mistral, and Llama families. The generality of our unsupervised learning method lends itself to applicability in a wide range of domains where external supervision is unavailable.
URL: https://openreview.net/forum?id=gInznr8EsQ
---
Title: Multitask Transformer Models for Demographic and Industry Profiling on Long-Form Blog Texts
Authors: Bahor Eshmirzayeva, Shirali Kadyrov
Abstract: We address the challenge of multitask author profiling on long-form blog text, where standard Transformer models are limited by fixed input length constraints that prevent full utilization of extended documents. To overcome this, we develop four transformer-based models that jointly predict gender, age group, and industry. Using a cleaned version of the Blog Authorship Corpus, we explore document-length handling strategies that span input ranges from 192 to 500 tokens, including long-context encoding, BART-based summarization, and chunked processing with prediction fusion. Our experiments show that multitask learning consistently outperforms strong single-task baselines, with the largest gains for industry. We further find that broader input context yields more reliable predictions, while alternative representations emphasize complementary stylistic and topical cues. Taken together, these findings provide a comprehensive analysis of text-length effects in multitask author profiling and highlight the importance of contextual breadth for robust demographic inference. The dataset was preprocessed by merging industry tags into fourteen categories and applying standard text normalization.
URL: https://openreview.net/forum?id=WtFwcCvt9i
---
Title: Differentially Private XGBoost Revisited: Is Random Decision Trees Really Better than Greedy Ones?
Authors: Erchi Wang, Arinbjörn Kolbeinsson, Luca Foschini, Yu-Xiang Wang
Abstract: Boosted Decision Trees (e.g., XGBoost) are one of the strongest and most widely used machine learning models. Motivated by applications in sensitive domains, various versions of Boosted Decision Tree learners with provably differential privacy (DP) guarantees were designed. Contrary to their non-private counterparts, a recent study shows that private boosting random decision trees outperform a more faithful privatization of XGBoost that uses greedy decision trees. In this paper, we challenge this conclusion with an improved DP-XGBoost algorithm and a thorough empirical study. Our results indicate that, although random selection remains strong on most datasets, our improved DP analysis narrows down the performance gap between random and greedy selection. At the same time, we identify regimes in which greedy selection clearly outperforms: when the number of trees is restricted to be small (e.g., for interpretability) or when interaction terms play a key role in prediction, random selection can suffer degradation in accuracy, whereas our DP-XGB greedy strategy remains robust and achieves better performance.
URL: https://openreview.net/forum?id=9fRDcavm3J
---
Title: Beyond Subtokens: A Rich Character Embedding for Low-Resource and Morphologically Complex Languages
Authors: Felix Schneider, Maria Gogolev, Sven Sickert, Joachim Denzler
Abstract: Tokenization and sub-tokenization based models like word2vec, BERT and the GPTs are
the state-of-the-art in natural language processing. Typically, these approaches have limi-
tations with respect to their input representation. They fail to fully capture orthographic
similarities and morphological variations, especially in highly inflected and under-resource
languages. To mitigate this problem, we propose to computes word vectors directly from
character strings, integrating both semantic and syntactic information. We denote this
transformer-based approach Rich Character Embeddings (RCE). Furthermore, we propose
a hybrid model that combines transformer and recurrent mechanisms. Both vector repre-
sentations can be used as a drop-in replacement for dictionary- and subtoken-based word
embeddings in existing model architectures. It has the potential to improve performance for
both large context-based language models like BERT and small models like word2vec for
under-resourced and morphologically rich languages. We evaluate our approach on various
tasks like SWAG, declension prediction for inflected languages, and metaphor and chias-
mus detection for various languages. Our experiments show that it outperforms traditional
token-based approaches on limited data using OddOneOut and TopK metrics.
URL: https://openreview.net/forum?id=4n4db5qmXZ
---
Title: On The Scalability Of Forward Gradients, Evolution Strategies, And Control Variates
Authors: Jake Levi, Seth Nabarro, Mark van der Wilk
Abstract: Stochastic gradient estimation methods such as Forward Gradients (FG) and Evolution Strategies (ES) have been proposed to overcome drawbacks of computing gradients with backpropagation (BP). However, FG and ES have large variance in high dimensions, connections between these methods have previously remained unclear, and while pure FG is guaranteed to be unbiased, proposed improvements have typically abandoned this property. We illuminate connections between FG and a popular variant of ES by proving mathematical equivalence on all quadratic objective functions. On an illustrative problem, we demonstrate theoretically how optimal convergence and learning rates scale unfavourably with intrinsic dimensionality and population size. We show that popular gradient descent techniques such as momentum and Adam do not address these fundamental scalability problems. We explore using control variates to reduce variance of FG while maintaining unbiasedness, and while we find limited success in improving over baselines, we also identify challenges that need to be overcome for these methods to scale effectively. Lastly we consider a biased method for variance reduction, and on a particular problem we show that this significantly outperforms the unbiased variance reduction methods that we consider. Assuming access to an asymptotically unbiased control variate, our results suggest that maintaining unbiasedness is not necessarily advantageous for variance reduction techniques, however we leave open the possibility that unbiasedness may be helpful when the control variate is asymptotically biased.
Code: https://github.com/jakelevi1996/forward_grad_public
URL: https://openreview.net/forum?id=s6g8yZimHE
---
Title: Reduced-Rank Outcome Compression for Causal Policy Optimization
Authors: Ezinne Nwankwo, Michael I. Jordan, Angela Zhou
Abstract: Evaluating the causal impacts of possible interventions is crucial for informing decision-making, especially towards improving access to opportunity. If causal effects are heterogeneous and predictable from covariates, then personalized treatment decisions can improve individual outcomes and contribute to both efficiency and equity. In practice, however, causal researchers do not have a single outcome in mind a priori and often collect multiple outcomes of interest that are noisy estimates of the true target of interest. For example, in government-assisted social benefit programs, policymakers collect many outcomes to understand the multidimensional nature of poverty. The ultimate goal is to learn an optimal treatment policy that in some sense maximizes multiple outcomes simultaneously. To address such issues, we present a data-driven dimensionality-reduction methodology for multiple outcomes in the context of optimal policy learning with multiple objectives. We learn a low-dimensional representation of the true outcome from the observed outcomes using reduced rank regression. We develop a suite of estimates that use the model to denoise observed outcomes, including commonly-used index weightings. These methods improve estimation error in policy evaluation and optimization, including on a case study of real-world cash transfer and social intervention data. Reducing the variance of noisy social outcomes can improve the performance of algorithmic allocations.
URL: https://openreview.net/forum?id=WQhOaY4yPC
---
Title: O$^2$-Searcher: A Searching-based Agent Model for Open-Domain Open-Ended Question Answering
Authors: Jianbiao Mei, Tao Hu, Daocheng Fu, Licheng Wen, Xuemeng Yang, Rong Wu, Pinlong Cai, Xinyu Cai, Xing Gao, Yu Yang, Chengjun Xie, Botian Shi, Yong Liu, Yu Qiao
Abstract: Large Language Models (LLMs), despite their advancements, are fundamentally limited by their static parametric knowledge, hindering performance on tasks requiring open-domain up-to-date information. While enabling LLMs to interact with external knowledge environments is a promising solution, current efforts primarily address closed-end problems. Open-ended questions, which characterized by lacking a standard answer or providing non-unique and diverse answers, remain underexplored. To bridge this gap, we present O2-Searcher, a novel search agent leveraging reinforcement learning to effectively tackle both open-ended and closed-ended questions in the open domain. O2-Searcher leverages an efficient, locally simulated search environment for dynamic knowledge acquisition, effectively decoupling the external world knowledge from model's sophisticated reasoning processes. It employs a unified training mechanism with meticulously designed reward functions, enabling the agent to identify problem types and adapt different answer generation strategies. Furthermore, to evaluate performance on complex open-ended tasks, we construct O2-QA, a high-quality benchmark featuring 300 manually curated, multi-domain open-ended questions with associated web page caches. Extensive experiments show that O2-Searcher, using only a 3B model, significantly surpasses leading LLM agents on O2-QA. It also achieves SOTA results on various closed-ended QA benchmarks against similarly-sized models, while performing on par with much larger ones.
URL: https://openreview.net/forum?id=rbIKFKFEeU
---
Title: Temporal Variational Implicit Neural Representations
Authors: Batuhan Koyuncu, Rachael DeVries, Ole Winther, Isabel Valera
Abstract: We introduce Temporal Variational Implicit Neural Representations (TV-INRs), a probabilistic framework for modeling irregular multivariate time series that enables efficient and accurate individualized imputation and forecasting. By integrating implicit neural representations with latent variable models, TV-INRs learn distributions over time-continuous generator functions conditioned on signal-specific covariates.
Unlike existing INR approaches that require extensive training, fine-tuning or meta-learning, our method achieves accurate individualized predictions through a single forward pass. Our experiments demonstrate that with a single TV-INRs instance, we can accurately solve diverse imputation and forecasting tasks, offering a computationally efficient and scalable solution for real-world applications. TV-INRs performs particularly well in low-data regimes, where on several datasets it achieves substantially lower imputation error, including order-of-magnitude improvements.
URL: https://openreview.net/forum?id=1CGfvw4ySe
---
Title: Data Selection Through Iterative Self-Filtering for Vision-Language Settings
Authors: Andrei Liviu Nicolicioiu, Sarvjeet Singh Ghotra, Morgane M Moss, Aaron Courville
Abstract: The availability of large amounts of clean data is paramount to training neural networks. However, at large scales, manual oversight is impractical, resulting in sizeable datasets that can be very noisy. Attempts to mitigate this obstacle to producing performant vision-language models have so far involved heuristics, curated reference datasets, and using pre-trained models. Here we propose a novel, bootstrapped method in which a CLIP model is trained on an evolving, self-selected dataset. This evolving dataset constitutes a balance of filtered, highly probable clean samples as well as diverse samples from the entire distribution. Our proposed Self-Filtering method iterates between training the model and selecting a subsequently improved data mixture. Training on vision-language datasets filtered by the proposed approach improves downstream performance without the need for additional data or pre-trained models.
URL: https://openreview.net/forum?id=F09NfCXuCe
---
Title: TERC: A Transfer Entropy Redundancy Criterion for State Variable Selection in Reinforcement Learning
Authors: Charles Westphal, Stephen Hailes, Mirco Musolesi
Abstract: Identifying the most suitable variables to represent the state is a fundamental challenge in Reinforcement Learning (RL). These variables must efficiently capture the information necessary for making optimal decisions. In order to address this problem, in this paper, we introduce the Transfer Entropy Redundancy Criterion (TERC), an information-theoretic criterion, which determines if there is \textit{entropy transferred} from observable state variables to actions during training. We define an algorithm based on TERC that provably excludes variables from the observable state that do not affect the agent's policy during learning. This yields compact state representations that reduce inference time by up to $2.6\times$.
Our approach is policy-dependent, making it agnostic to the underlying learning algorithm. The efficiency gains we demonstrate arise at retraining and inference time on the reduced state.
Our method improves both retraining and inference efficiency. We demonstrate its effectiveness across three distinct algorithm classes, namely tabular Q-learning, Actor-Critic, and Proximal Policy Optimization (PPO), evaluated in a range of environments. Furthermore, to highlight the differences between the proposed methodology and the current state-of-the-art feature selection approaches, we present a series of controlled experiments on synthetic data, before generalizing to real-world decision-making tasks. We also introduce a representation of the problem that compactly captures the transfer of information from observable state variables to actions as Bayesian networks.
URL: https://openreview.net/forum?id=J0ad21E0vX
---
Title: Judging the Judges: A Systematic Evaluation of Bias Mitigation Strategies in LLM-as-a-Judge Pipelines
Authors: Sadman Kabir Soumik
Abstract: LLM-as-a-Judge has become the dominant paradigm for evaluating language model outputs, yet LLM judges exhibit systematic biases that compromise evaluation reliability. We present a comprehensive empirical study comparing nine debiasing strategies across five judge models from four provider families (Google, Anthropic, OpenAI, Meta), three benchmarks (MT-Bench n=400, LLMBar n=200, custom n=375), and four bias types. Our headline practical finding is that a mid-tier model with the right debiasing can outperform frontier judges at a fraction of the cost: Gemini 2.5 Flash with the Combined Budget strategy achieves the highest agreement of any configuration we tested (71.0%, κ = 0.549, p < 0.0001) at ~$0.001 per evaluation, roughly 15× cheaper than the strongest frontier configuration (Claude Sonnet 4 with the same strategy at 69.5%, ~$0.015 per evaluation). Our other key findings: (1) Style bias is the dominant bias (0.10–0.76 baseline across all models, mostly favoring markdown over plain prose), far exceeding position bias (≤ 0.04), yet has received minimal research attention. (2) Verbosity bias is heterogeneous across models when measured length-aware: Llama, Gemini Pro, and Gemini Flash show classical verbosity bias (+0.24 to +0.44, prefer longer), Claude Sonnet 4 shows the opposite (-0.12, prefer concise), and GPT-4o is essentially neutral (-0.04); on truncation controls all models correctly prefer the genuinely complete response (0.88–1.00 accuracy), so the expansion-pair preferences cannot be reduced to length-only effects. (3) Debiasing is statistically beneficial for multiple models: Claude S8 (+11.5 pp, p<0.0001), Flash S8 (+7.5 pp, p<0.0001), Claude S5 (+7.3 pp, p=0.0009) survive Holm-Bonferroni correction; Flash S1 (+4.7 pp, p=0.004) and Llama S8 (+4.5 pp, p=0.011) are significant before correction; Pro and GPT-4o show smaller, non-significant directional gains. We release our evaluation framework, the 375-pair controlled dataset (now including round-robin MODEL_ORIGIN and position-mirrored STYLE pairs), and per-instance cached results for all 9 strategies.
URL: https://openreview.net/forum?id=QF4lAmG4zc
---
Title: Adaptive Budget Allocation for Orthogonal Subspace Adapter Tuning in LLMs Continual Learning
Authors: Zhiyi Wan, Wanrou Du, Yijia Chi, Liang Li, Miao Pan, Xiaoqi Qin
Abstract: Large language models (LLMs) often suffer from catastrophic forgetting in continual learning (CL) scenarios, where performance on previously learned tasks degrades severely while training on sequentially arriving tasks.
Although pioneering CL approaches using orthogonal subspaces can mitigate task interference, they typically employ fixed budget allocation, neglecting the varying complexity across tasks and layers.
Besides, recent budget-adaptive tuning methods for LLMs often adopt multi-stage paradigms that decouple optimization and budget allocation. Such decoupling results in potential misalignment, which hinders those approaches' practical application in CL scenarios.
To address these limitations, we propose OA-Adapter, a novel parameter-efficient approach for continual learning in LLMs that unifies dynamic budget adaptation with orthogonal subspace learning in an end-to-end training stage.
Specifically, OA-Adapter introduces a dynamic bottleneck dimension adaptation mechanism that simultaneously allocates an efficient parameter budget and optimizes task objectives without misalignment.
To effectively preserve previously acquired knowledge while coordinating with the dynamic budget allocation, orthogonal constraints are applied specifically between the parameter subspace of the current task and the dynamically allocated parameter subspaces of historical tasks.
Experimental results on continual learning benchmarks demonstrate that OA-Adapter outperforms state-of-the-art methods in both accuracy and parameter efficiency. OA-Adapter achieves higher average accuracy while using \(58.5\%\) fewer parameters on the Standard CL Benchmark, and maintains its advantages on two larger benchmarks comprising 15 tasks.
URL: https://openreview.net/forum?id=LNGDlLOdex
---
Title: TinyV: Reducing False Negatives in Verification Improves RL for LLM Reasoning
Authors: Zhangchen Xu, Yuetai Li, Fengqing Jiang, Bhaskar Ramasubramanian, Luyao Niu, Basel Alomair, Bill Yuchen Lin, Radha Poovendran
Abstract: Reinforcement Learning (RL) has become a powerful tool for enhancing the reasoning abilities of large language models (LLMs) by optimizing their policies with reward signals. Yet, RL’s success relies on the reliability of rewards, which are provided by verifiers. In this paper, we expose and analyze a widespread problem—false negatives—where verifiers incorrectly reject correct model outputs. Our in-depth study of the Big-Math-RL-Verified dataset reveals that over 38% of model-generated responses suffer from false negatives in math reasoning tasks, where the verifier fails to recognize correct answers. We show, both empirically and theoretically, that these false negatives severely impair RL training by depriving the model of informative gradient signals and slowing convergence. To mitigate this, we propose TinyV, a lightweight LLM-based verifier that augments rule-based methods with a compact LLM to dynamically detect false negatives and recover valid trajectories in rule-based final-answer verification settings. Across multiple math-reasoning benchmarks, integrating TinyV improves final model performance by up to 10%, and reaches the peak performance of the rule-based verifier using fewer than 50% of the training steps. Our findings highlight the critical importance of addressing verifier false negatives in RL for verifiable reasoning tasks, particularly in math domains. Our code is available at: https://github.com/uw-nsl/TinyV.
URL: https://openreview.net/forum?id=HMGsqApBM3
---
Title: Firewalls to Secure Dynamic LLM Agentic Networks
Authors: Sahar Abdelnabi, Amr Gomaa, Eugene Bagdasarian, Per Ola Kristensson, Reza Shokri
Abstract: The emergence of agent-to-agent communication protocols mirrors the early internet: powerful connectivity with minimal security infrastructure. When AI agents communicate on behalf of users, every message crosses a trust boundary where the user's personal data and the external agent's unconstrained language each present distinct risks. We address both through a dual-firewall architecture grounded in a unifying principle: each task defines a context, and both sides of the communication carry information far exceeding what that context requires. Our firewalls act as projections onto the task context, allowing only contextually appropriate content to cross each boundary. The Language Converter Firewall projects incoming messages onto a closed, domain-specific, structured protocol; an external agent's message is converted to validated fields while persuasive framing, urgency tactics, and embedded instructions are structurally eliminated through deterministic verification. This replaces the asymmetric challenge of resisting every possible manipulation with the structural guarantee that manipulation has no channel through which to arrive. The Data Abstraction Firewall projects outgoing information onto the granularity appropriate for the task, rather than applying binary disclose-or-redact filtering, as previous airgapping solutions did. Both firewalls operate in a trusted environment isolated from external input, applying domain-specific rules learned automatically from demonstrations. Across 864 attacks spanning three domains on the recent ConVerse benchmark, our architecture reduces privacy attack success rates (e.g., from 84% to 10% for GPT-5) and security attacks (from 60% to 3%), while maintaining or even improving task completion quality.
URL: https://openreview.net/forum?id=w02FW1dMoY
---
Title: DAG-aware GAT for Causal Effect Estimation
Authors: Manqing Liu, David Remy Bellamy, Andrew Beam
Abstract: Causal inference from observational data is a cornerstone of decision-making in healthcare, economics, and the social sciences. While deep learning has significantly advanced effect estimation, standard architectures often fail to respect the structural constraints inherent in causal systems, leading to biased results in complex scenarios like proximal inference. In this paper, we introduce the \textbf{DAG-aware Graph Attention Network (GAT)}, a novel neural framework that bridges structural causal modeling with graph representation learning. Unlike traditional Transformers or unconstrained GNNs, our model embeds the causal Directed Acyclic Graph (DAG) as a hard structural inductive bias directly into the attention mechanism. This ensures that information flow strictly adheres to valid causal pathways while preserving the semantic integrity of heterogeneous variables by omitting distortive normalization layers. Extensive experiments on several benchmark datasets show that the DAG-aware GAT consistently outperforms classical non-parametric baselines, modular MLPs, and causally-agnostic graph architectures. By prioritizing causal integrity over generic predictive heuristics, our approach provides a robust and interpretable foundation for reasoning from complex observational data. Our implementation is available at \url{https://github.com/ManqingLiu/DAG-aware-GAT}.
URL: https://openreview.net/forum?id=pzFwkjRVXy
---
Title: NAS Without Priors: A Robust Architecture Search Framework for Unseen-Data
Authors: Shahid Siddiqui, Christos Kyrkou, Theocharis Theocharides
Abstract: Neural architecture search (NAS) has been widely used to automate neural network design for image classification; however, most NAS research has focused on CIFAR, ImageNet, and their derivative benchmarks. These datasets benefit from well-established architecture design practices, data preprocessing techniques, and training protocols developed prior to NAS, causing many NAS methods to meta-overfit and struggle to generalize to entirely novel datasets. In this work, we analyze the limitations of existing NAS practices and propose a framework specifically designed to generalize to unseen-data. In contrast to the prevailing paradigm of exploring extremely large search spaces using low-fidelity evaluations, we advocate sparser exploration combined with high-fidelity performance estimation. We demonstrate that macro-architecture variations alone induce substantial architectural diversity, and that concentrating computational resources on high-fidelity evaluation of fewer candidates produces reliable reward signals enabling better architecture discovery. To obtain robust candidate rankings, we repeatedly train architectures on the entire training set using multiple random seeds. While this approach substantially reduces performance variance due to random seed variability and enables accurate candidate ranking, it comes at a significant computational cost. To mitigate the cost of such high-fidelity evaluation, particularly for larger or high-resolution datasets, we introduce a dataset- and architecture-aware multi-fidelity search strategy that both reduces computational overhead and stabilizes candidate rankings under varying fidelity levels. We evaluate our framework on the NAS Unseen-Data Challenge (https://www.nascompetition.com/archive), where, under an 8-hour budget per dataset (including search and training), it ranks first with a combined score of 12.19, outperforming both manually designed baselines and the second- and third-place NAS solutions (10.89 and 10.43). Moreover, we test our framework on CIFAR-10 to enable comparison with the broader NAS literature, despite our primary focus on NAS generalization beyond conventional benchmarks. Our framework achieves competitive error rates against a broad spectrum of macro- and micro-NAS methods, demonstrating the effectiveness of sparse, high-fidelity evaluations and the proposed multi-fidelity search algorithm. Our code and search logs are available at: (https://github.com/siddikui/NAS-Challenge-2025-Solution).
URL: https://openreview.net/forum?id=PaEk1gYrFz
---
Title: Task Vector Bases: A Unified and Scalable Framework for Compressed Task Arithmetic
Authors: Siqi Zeng, Yifei He, Meitong Liu, Weiqiu You, Yifan HAO, Yao-Hung Hubert Tsai, Makoto Yamada, Han Zhao
Abstract: Task arithmetic, representing downstream tasks through linear operations on task vectors, has emerged as a simple yet powerful paradigm for transferring knowledge across diverse settings. However, maintaining a large collection of task vectors introduces scalability challenges in both storage and computation. We propose Task Vector Bases, a framework compressing $T$ task vectors into $M < T$ basis vectors while preserving the functionality of task arithmetic. By representing each task vector as a structured linear combination of basis atoms, our approach supports standard operations such as addition, negation, as well as more advanced arithmetic ones. The framework is orthogonal to other efficiency-oriented improvements in task arithmetic and can be used in combination with them. We provide theoretical analysis showing that basis compression retains addition generalization guarantees and enables principled unlearning, with error bounds depending on reconstruction quality. Empirically, our proposed basis construction methods consistently outperform heuristic basis construction baselines and, in some cases, even surpass the performance of full task vector collections across diverse downstream applications while reducing storage and computational requirements.
URL: https://openreview.net/forum?id=zkc7u3mIaE
---
Title: CentroidKV: Efficient Long-Context LLM Inference via KV Cache Clustering
Authors: Jie Hu, Shengnan Wang, Yutong He, Ping Gong, Jiawei Yi, Juncheng Zhang, Youhui Bai, Renhai Chen, Gong Zhang, Cheng Li, Kun Yuan
Abstract: Large language models (LLMs) with extended context windows have become increasingly prevalent for tackling complex tasks. However, the substantial Key-Value (KV) cache required for long-context LLMs poses significant deployment challenges. Existing approaches either discard potentially critical information needed for future generations or offer limited efficiency gains due to high computational overhead. In this paper, we introduce \textit{CentroidKV}, a simple yet effective framework for online KV cache clustering. Our approach is based on the observation that key states exhibit high similarity along the sequence dimension. To enable efficient clustering, we divide the sequence into chunks and propose \textit{Chunked Soft Matching}, which employs an alternating partition strategy within each chunk and identifies clusters based on similarity. CentroidKV then merges the KV cache within each cluster into a single centroid. Additionally, we provide a theoretical analysis of the computational complexity and the optimality of the intra-chunk partitioning strategy. Extensive experiments across various models and long-context benchmarks demonstrate that CentroidKV achieves up to 75\% reduction in KV cache memory usage while maintaining comparable model performance. Moreover, with minimal computational overhead, CentroidKV accelerates the decoding stage of inference by up to 1.92$\times$ and increases the serving throughput by up to 4$\times$.
URL: https://openreview.net/forum?id=T3EeupQhGj
---
Title: PERRY: Policy Evaluation with Confidence Intervals using Auxiliary Data
Authors: Aishwarya Mandyam, Jason Meng, Ge Gao, Jiankai Sun, Mac Schwager, Barbara E Engelhardt, Emma Brunskill
Abstract: Off-policy evaluation (OPE) methods estimate the value of a new reinforcement learning (RL) policy prior to deployment. Recent advances have shown that leveraging auxiliary datasets, such as those synthesized by generative models, can improve the accuracy of OPE methods. Unfortunately, such auxiliary datasets may also be biased, and existing methods for using data augmentation within OPE lack principled uncertainty quantification. In high stakes domains like healthcare, reliable uncertainty estimates are important for ensuring safe and informed deployment of RL policies. In this work, we propose two methods to construct valid confidence intervals for OPE with data augmentation. The first provides a confidence interval over $V^{\pi}(s)$, the policy value conditioned on an initial state $s$. To do so we introduce a new conformal prediction method suitable for Markov Decision Processes (MDPs) with continuous state spaces, extending prior work to higher-dimensional settings. Second, we consider the more common task of estimating the average policy performance over all initial states, $V^{\pi}$; we introduce a method that draws on ideas from doubly robust estimation and prediction powered inference. Across simulators spanning inventory management, robotics, healthcare, and a real healthcare dataset from MIMIC-IV, we find that our methods can effectively leverage auxiliary data and consistently produce confidence intervals that cover the ground truth policy values, unlike previously proposed methods. Our work enables a future in which OPE can provide rigorous uncertainty estimates for high-stakes domains.
URL: https://openreview.net/forum?id=RbecZhn2qX
---
Title: LoLA: Low-Rank Linear Attention with Sparse Caching
Authors: Luke McDermott, Robert W. Heath, Rahul Parhi
Abstract: The per-token cost of transformer inference scales with context length, preventing its application to lifelong in-context learning. Linear attention is an efficient alternative that maintains a constant memory footprint, even on infinite context lengths. While this is a potential candidate for lifelong learning, it falls short in memory capacity. In this paper, we propose LoLA, a training-free augmentation to linear attention that boosts associative recall. LoLA distributes past key-value pairs from context into three memory systems: (i) recent pairs in a local sliding window cache; (ii) difficult-to-memorize pairs in a sparse, global cache; and (iii) generic pairs in the recurrent hidden state of linear attention. We show through ablations that our self-recall error metric is crucial to efficiently manage long-term associative memories. On pass-key retrieval tasks, LoLA improves the base model's performance from 0.6% to 97.4% accuracy. This is achieved with a 4.6x smaller cache than Llama-3.1 8B on 4K context length. LoLA also outperforms other 1B and 8B parameter subquadratic models on zero-shot commonsense reasoning tasks.
URL: https://openreview.net/forum?id=3KhDA252y3
---
Title: Gaga: Group Any Gausians via 3D-aware Memory Bank
Authors: Weijie Lyu, Xueting Li, Abhijit Kundu, Yi-Hsuan Tsai, Ming-Hsuan Yang
Abstract: We introduce Gaga, a framework that reconstructs and segments open-world 3D scenes by leveraging inconsistent 2D masks predicted by zero-shot class-agnostic segmentation models. Contrasted to prior 3D scene segmentation approaches that rely on video object tracking
or contrastive learning methods, Gaga utilizes spatial information and effectively associates object masks across diverse camera poses through a novel 3D-aware memory bank. By eliminating the assumption of continuous view changes in training images, Gaga demonstrates robustness to variations in camera poses, particularly beneficial for sparsely sampled images, ensuring precise mask label consistency. Furthermore, Gaga accommodates 2D segmentation masks from diverse sources and demonstrates robust performance with different open-world zero-shot class-agnostic segmentation models, significantly enhancing its versatility. Extensive qualitative and quantitative evaluations demonstrate that Gaga performs favorably against state-of-the-art methods, emphasizing its potential for real-world applications such as 3D scene understanding and manipulation.
URL: https://openreview.net/forum?id=cC1TLyK3iW
---
Title: Trace Reconstruction with Language Models
Authors: Franziska Weindel, Michael Girsch, Reinhard Heckel
Abstract: The general trace reconstruction problem seeks to recover an original sequence from its noisy copies independently corrupted by insertions, deletions, and substitutions. This problem arises in applications such as DNA data storage, a promising storage medium due to its high information density and longevity. However, errors introduced during DNA synthesis, storage, and sequencing require correction through algorithms and codes, with trace reconstruction often used as part of data retrieval. In this work, we propose TReconLM, a decoder-only transformer that solves trace reconstruction as a next-token prediction task. TReconLM outperforms state-of-the-art trace reconstruction algorithms, including prior deep-learning approaches, recovering a substantially higher fraction of sequences without error. We pretrain on synthetic data generated from a simple error model and fine-tune on real-world data to adapt to technology-specific error patterns. Code is available at \url{https://github.com/MLI-lab/TReconLM}.
URL: https://openreview.net/forum?id=k2zAyJAfnj
---
Title: Rethinking On-policy Optimization for Query Augmentation
Authors: Zhichao Xu, Shengyao Zhuang, Xueguang Ma, Bingsen Chen, Yijun Tian, Fengran Mo, Tao Li, Jie Cao, Vivek Srikumar
Abstract: Recent advances in large language models (LLMs) have led to a surge of interest in query augmentation for information retrieval (IR). Two main approaches have emerged. The first prompts LLMs to generate answers or pseudo-documents that serve as new queries, relying purely on the model's parametric knowledge or contextual information. The second applies reinforcement learning (RL) to fine-tune LLMs for query rewriting, directly optimizing retrieval metrics. While having respective advantages and limitations, the two approaches have not been compared under consistent experimental conditions. In this work, we present the first systematic comparison of prompting-based and RL-based query augmentation across diverse benchmarks, including evidence-seeking, ad hoc, and tool retrieval. Our key finding is that *under a compute-aware comparison setting, simple, training-free query augmentation often performs on par with, or even surpasses, more expensive RL-based counterparts, especially when using powerful LLMs*. Motivated by this discovery, we introduce a novel hybrid method, On-policy Pseudo-document Query Expansion (OPQE), in which the LLM policy learns to generate a pseudo-document that maximizes retrieval performance, rather than rewriting the query, thus merging the flexibility and generative structure of prompting with the targeted optimization of RL. We show OPQE outperforms both standalone prompting and RL-based rewriting, demonstrating that a synergistic approach yields the best results. We open source our implementation to facilitate reproducibility.
URL: https://openreview.net/forum?id=mmqbjhz5Br
---
Title: Beyond the Linear Separability Ceiling: Aligning Representations in VLMs
Authors: Enrico Vompa, Tanel Tammet, Mohit Vaishnav
Abstract: A challenge in advancing Visual-Language Models (VLMs) is determining whether their failures on abstract reasoning tasks, such as Bongard problems, stem from flawed perception or faulty top-down reasoning. To disentangle these factors, we introduce a diagnostic framework centered on the Linear Separability Ceiling (LSC), the performance achievable by a linear classifier on a VLM's raw visual embeddings. Applying this framework to state-of-the-art VLMs, we uncover a pervasive ''alignment gap'', where most models fail to generatively outperform the linear separability of their representations. We find that the few models surpassing this ceiling do so via two mechanisms: by further refining visual representations into a more linearly separable format or by executing non-linear decision logic. We demonstrate that this bottleneck is not a fundamental limitation but a solvable visual alignment issue. Our method augments standard next-token prediction with a contrastive objective to restructure the visual manifold into a more one-dimensionally linear geometry, improving image-to-image comparison and enabling models to significantly surpass the LSC on abstract compositional reasoning tasks.
URL: https://openreview.net/forum?id=3uX4p80bN0
---
Title: PLGC: Pseudo-Labeled Graph Condensation
Authors: Jay Nandy, Arnab Kumar Mondal, Anuj Rathore, Mahesh Chandran
Abstract: Large graph datasets make training graph neural networks (GNNs) computationally costly. Graph condensation methods address this by generating small synthetic graphs that approximate the original data. However, existing approaches rely on clean, supervised labels, which limits their reliability when labels are scarce, noisy, or inconsistent. We propose Pseudo-Labeled Graph Condensation (PLGC), a self-supervised framework that constructs latent pseudo-labels from node embeddings and optimizes condensed graphs to match the original graph’s structural and feature statistics-without requiring ground-truth labels. PLGC offers three key contributions: (1) A diagnosis of why supervised condensation fails under label noise and distribution shift. (2) A label-free condensation method that jointly learns latent prototypes and node assignments. (3) Theoretical guarantees showing that pseudo-labels preserve latent structural statistics of the original graph and ensure accurate embedding alignment. Empirically, across node classification and link prediction tasks, PLGC achieves competitive performance with state-of-the-art supervised condensation methods on clean datasets and exhibits substantial robustness under label noise, often outperforming all baselines by a significant margin. Our findings highlight the practical and theoretical advantages of self-supervised graph condensation in noisy or weakly-labeled environments.
URL: https://openreview.net/forum?id=TkpewrzsnJ
---
Title: Uncertainty-Aware Safety Propagation Critics for Safe Reinforcement Learning
Authors: Kutay Demiray, Efe Eren Ceyani, Ozgur S. Oguz
Abstract: Safe reinforcement learning (RL) aims to optimize long-term performance while satisfying safety constraints, a requirement that is critical in many applications but difficult to guarantee when cost estimates are inaccurate or data is limited. In model-free actor-critic methods, cost critics are often unreliable in poorly explored regions, leading to constraint violations during both training and deployment. In this work, we propose a novel uncertainty-aware approach in safe RL called USPC, which constructs conservative cost surrogates using epistemic uncertainty. Our method trains an ensemble of cost critics to estimate uncertainty and uses these estimates to build an upper confidence bound on predicted costs. We then introduce a safe set network that approximates a pessimistic surrogate of the cost action-value function inspired by safe Bayesian optimization, enabling scalable safety propagation in continuous state-action spaces. Replacing standard cost critics with this surrogate in existing off-policy safe RL algorithms yields policies that are less likely to violate cost constraints. We show empirically across multiple Safety Gymnasium benchmark tasks that our approach reduces both the frequency and magnitude of constraint violations in most tasks while maintaining competitive reward performance compared to the baselines. We further provide theoretical justification for our approach, including a conservativeness guarantee for the safe set network, and an approximation result for the anchor-based scheme.
URL: https://openreview.net/forum?id=b8GWqgLEHh
---
Title: Predictive Feature Caching for Training-free Acceleration of Molecular Geometry Generation
Authors: Johanna Sommer, Nils Fleischmann, John Rachwan, Stephan Günnemann, Bertrand Charpentier
Abstract: Flow matching models generate high-fidelity molecular geometries but incur significant computational costs during inference, requiring hundreds of neural network evaluations. This inference cost becomes the primary bottleneck when such models are employed in practice to sample large numbers of molecular candidates. This work presents a training-free caching strategy that accelerates molecular geometry generation by predicting intermediate hidden states across solver steps. This caching scheme operates directly on the SE(3)-equivariant
backbone, is compatible with pretrained models, and is orthogonal to existing training-based accelerations and system-level optimizations. Experiments on molecular geometry generation demonstrate that caching achieves a twofold reduction in wall-clock inference time at matched sample quality and a speedup of up to 3× with minimal sample quality degradation. Because these gains compound with other optimizations, applying caching alongside other general, lossless optimizations yield as much as a 7× speedup.
URL: https://openreview.net/forum?id=NaLVutCHCI
---
Title: Inference-Time Scaling for Joint Audio-Video Generation
Authors: Jaemin Jung, Kyeongha Rho, Inkyu Shin, Joon Son Chung
Abstract: Joint audio-video generation aims to synthesize realistic audio-video pairs that are both semantically aligned with text prompts and precisely synchronized. While existing joint audio-video generation models often require substantial training resources to improve fidelity, Inference-Time Scaling (ITS) has recently emerged as a promising training-free alternative in single-modality domains. However, extending ITS from a single modality to multimodal domains is non-trivial, as it requires balancing multiple heterogeneous objectives. In this paper, we present the first comprehensive study of ITS for joint audio-video generation. We first demonstrate that a multi-verifier framework is essential to address the limitations of single-objective guidance, including asymmetric performance trade-offs and verifier hacking. Through systematic analysis, we then identify an optimal multi-verifier combination that yields balanced improvements across all quality dimensions. Finally, to effectively aggregate diverse reward signals, we propose Adaptive Reward Weighting (ARW), a novel test-time optimization algorithm. ARW treats reward aggregation as an online optimization problem, utilizing learnable parameters to calibrate reward variances without requiring prior knowledge of reward distributions, thereby ensuring robust multi-objective selection. Experimental results on VGGSound and JavisBench-mini benchmarks demonstrate that our framework significantly enhances semantic alignment, perceptual quality, and audio-visual synchronization of generated outputs.
URL: https://openreview.net/forum?id=MHNFjjm5nO
---
Title: Zero-Shot Model Search via Text-to-Logit Matching
Authors: Jonathan Kahana, Or Nathan, Eliahu Horwitz, Yedid Hoshen
Abstract: With the increasing number of publicly available models, there are pre-trained, online models for many tasks that users require. In practice, users cannot find the relevant models as current search methods are text-based using the documentation which most models lack of. This paper presents ProbeLog, a method for retrieving classification models that can recognize a target concept, such as "Dog", without access to model metadata or training data. Specifically, ProbeLog computes a descriptor for each output dimension (logit) of each model, by observing its responses to a fixed set of inputs (probes). Similarly, we compute how the target concept is related to each probe. By measuring the distance between the probe responses of logits and concepts, we can identify logits that recognize the target concept. This enables zero-shot, text-based model retrieval ("find all logits corresponding to dogs"). To prevent hubbing, we calibrate the distances of each logit, according to other closely related concepts. We demonstrate that ProbeLog achieves high retrieval accuracy, both in ImageNet and real-world fine-grained search tasks, while being scalable to full-size repositories. Importantly, further analysis reveals that the retrieval order is highly correlated with model and logit accuracies, thus allowing ProbeLog to find suitable and accurate models for users tasks in a zero-shot manner.
URL: https://openreview.net/forum?id=m4Qmst7iH7
---
Title: ImageNot: A contrast with ImageNet preserves model rankings
Authors: Olawale Elijah Salaudeen, Moritz Hardt
Abstract: We introduce ImageNot, a dataset constructed explicitly to be drastically different than ImageNet while matching its scale. ImageNot is designed to test the external validity of deep learning progress on ImageNet. We show that key model architectures developed for ImageNet over the years rank identically to how they rank on ImageNet when trained from scratch and evaluated on ImageNot. Moreover, the relative improvements of each model over earlier models strongly correlate in both datasets. Our work demonstrates a surprising degree of external validity in the relative performance of image classification models when trained and evaluated on an entirely different dataset. This stands in contrast with absolute accuracy numbers that typically drop sharply even under small changes to a dataset.
URL: https://openreview.net/forum?id=YVbhMerXv9
---
Title: An Efficient End-to-End Framework for Localized Second-Order Pooling
Authors: Zhongyan Zhang, Saimunur Rahman, Piotr Koniusz, Luping Zhou, Lei Wang
Abstract: Second-order pooling has proven effective for deep image classification by representing an image with the covariance matrix of its local feature descriptors. However, the diverse visual content in images leads to local descriptors being distributed as multiple modes in the feature space, limiting the effectiveness of a single global covariance matrix. In this work, we propose an efficient end-to-end framework for localized second-order pooling. Our approach jointly learns clusters of local feature descriptors across the entire training set, adaptively assigns the descriptors of each image to the appropriate clusters, computes a localized covariance matrix based on the descriptors assigned to each cluster, and then integrates these matrices to form the image representation.This is achieved through an attention-based local cluster mining branch that automatically identifies the clusters that each local descriptor shall be assigned to. Furthermore, to manage the significant computational overhead incurred by the use of multiple local covariance matrices, we utilize a simple but efficient sample-adaptive feature fusion scheme. This scheme adaptively generates fusion weights for each localized covariance matrix using a lightweight predictor, ensuring both computational efficiency and flexibility. Extensive experiments on multiple fine-grained and large-scale image classification datasets demonstrate that our method consistently improves performance when integrated with state-of-the-art second-order pooling methods and leading network architectures. Ablation studies further verify the efficiency of our feature fusion scheme compared to the existing common alternatives.
URL: https://openreview.net/forum?id=nJBeXjxAqx
---
Title: Fine-Grained Uncertainty Quantification for Long-Form Language Model Outputs: A Comparative Study
Authors: Dylan Bouchard, Mohit Singh Chauhan, Viren Bajaj, David Skarbrevik
Abstract: Uncertainty quantification has emerged as an effective approach to closed-book hallucination detection for LLMs, but existing methods are largely designed for short-form outputs and do not generalize well to long-form generation. We introduce a taxonomy for fine-grained uncertainty quantification in long-form LLM outputs that distinguishes methods by design choices at three stages: response decomposition, unit-level scoring, and response-level aggregation. We formalize several families of consistency-based black-box scorers, providing generalizations and extensions of existing methods. We also introduce FactScore-STEM-Geo, a new 400-question long-form QA dataset spanning four categories across STEM and Geography. In our experiments across multiple LLMs and datasets, we find 1) claim-response entailment consistently performs better or on par with more complex claim-level scorers, 2) claim-level scoring generally yields better results than sentence-level scoring, and 3) uncertainty-aware decoding is highly effective for improving the factuality of long-form outputs. Our framework clarifies relationships between prior methods, enables apples-to-apples comparisons, and provides practical guidance for selecting components for fine-grained UQ.
URL: https://openreview.net/forum?id=gngp4Zz9Sj
---
Title: Forge: Foundational Optimization Representations from Graph Embeddings
Authors: Zohair Shafi, Serdar Kadioglu
Abstract: Combinatorial optimization problems are ubiquitous in science and engineering. Still, learning-based approaches to accelerate combinatorial optimization often require solving a large number of difficult instances to collect training data, incurring significant computational cost. Existing learning-based methods require training dedicated models for each problem distribution, for each downstream task, severely limiting their scalability and generalization. We introduce Forge: Foundational Optimization Representations from Graph Embeddings, a framework that pre-trains a vector-quantized graph autoencoder on a large, diverse collection of mixed-integer programming (MIP) instances in an unsupervised manner, without relying on optimization solvers or optimal solutions. Vector quantization produces discrete code assignments that serve as a vocabulary for representing optimization instances. We evaluate Forge in both unsupervised and supervised settings. In the unsupervised setting, Forge embeddings effectively cluster unseen instances across problem domains and sizes. In the supervised setting, we fine-tune Forge embeddings and show that a single pre-trained model helps predicting both the integrality gap for cut-generation and variable hints for search guidance across multiple problem and size distributions. In both tasks, we improve the performance of a commercial optimization solver and outperform state-of-the-art learning-based methods. Finally, we open-source our training code, pre-trained Forge weights, and embeddings for multiple MIP distributions to foster further research in representation learning for optimization problems.
URL: https://openreview.net/forum?id=7Uo1yRWUpo
---
Title: Adapting to Any Bit-Width: Channel-Wise Mixed-Precision Quantization for LLMs
Authors: Zihan Chen, BIKE XIE, Jundong Li, Cong Shen
Abstract: Large Language Models (LLMs) have demonstrated remarkable success across a wide range of language tasks, but their deployment on edge devices remains challenging due to the substantial memory requirements imposed by their large parameter sizes. Weight-only quantization presents a promising solution to reduce the memory footprint of LLMs. However, existing approaches primarily focus on integer-bit quantization, limiting their adaptability to fractional-bit quantization tasks and preventing the full utilization of available storage space on devices. In this paper, we introduce Channel-Wise Mixed-Precision Quantization (CMPQ), a novel mixed-precision quantization method that allocates quantization precision in a channel-wise pattern based on activation distributions. By assigning different precision levels to different weight channels, CMPQ supports arbitrary average bit-widths in the low-bit regime (e.g., between 2 and 4 bits). CMPQ employs a non-uniform quantization strategy and incorporates two outlier extraction techniques that collaboratively preserve the critical information, thereby minimizing the quantization loss. Experiments on nine different LLMs demonstrate that CMPQ not only enhances performance in integer-bit quantization tasks but also achieves significant performance gains with a modest increase in memory usage by performing in a mixed-precision way. CMPQ represents an adaptive and effective approach to LLM quantization, offering substantial benefits across diverse device capabilities.
URL: https://openreview.net/forum?id=1t6sEhdLxf
---
Title: Reproducing FACTER: Fairness via Conformal Thresholding and Prompt Repair
Authors: Oscar Miró López-Feliu, Daimy van Loo, Xanthos Kekkos, Mikel Blom, Clara Rus
Abstract: Fayyazi et al. (2025) recently proposed FACTER, a model-agnostic framework designed to jointly enforce fairness and statistical coverage in LLM-based recommendation through conformal thresholding and iterative prompt repair. In this work, we conduct a reproducibility study of the FACTER framework across diverse architectures and dataset sparsity levels, evaluating both the original open-ended generation task and a constrained re-ranking extension. Under the strict reproduction, we observe a divergence in recommendation utility, which we trace to underspecified target-set evaluation in the original study. We then use the constrained re-ranking setting to evaluate FACTER when the candidate set is fixed, and introduce a static Fair Zero-Shot baseline to isolate the contribution of the iterative prompt repair loop. Our analysis shows that FACTER consistently reduces adaptive-threshold violation counts, but that these reductions are not consistently reflected under the fixed threshold or in global fairness metrics. In the constrained ranking setting, static fairness instructions achieve comparable semantic-parity outcomes to FACTER's dynamic repair loop, suggesting that the additional online repair mechanism provides limited benefit in this formulation. All code and reproduction artifacts are available at https://github.com/oscar-omlf/facter-repr.
URL: https://openreview.net/forum?id=4BPFVex4EM
---
Title: Fac-TDMPC: A Factored World Model for Robot Planning
Authors: Yuan Zhang, Jianhong Wang, Jinke He, Frans A Oliehoek, Joschka Boedecker
Abstract: Model-based reinforcement learning (MBRL) has shown strong sample efficiency in robotics by learning predictive world models and planning with them, but existing methods suffer from high planning latency due to the combination of centralized world models and model predictive control (MPC) as planners, thus limiting the real-time deployment in high-dimensional action spaces. We introduce \textbf{Fac-TDMPC}, a factored latent-space world model that decomposes transition, reward, and value functions on the latent space and learns the factorization via model distillation. The factored design enables decentralized planning across action dimensions. Empirically, Fac-TDMPC achieves substantial planning speedups while preserving the control performance across a suite of continuous-control robotic tasks; it also demonstrates improved robustness to action perturbations, interpretable joint-level latent structure, and enhanced multi-task data efficiency.
URL: https://openreview.net/forum?id=Smb0sdocmd
---
Title: Hybrid Combinatorial Multi-armed Bandits with Probabilistically Triggered Arms
Authors: Kongchang Zhou, Tingyu Zhang, Wei Chen, Fang Kong
Abstract: The problem of combinatorial multi-armed bandits with probabilistically triggered arms (CMAB-T) has been extensively studied. Prior work primarily focuses on either the online setting where an agent learns about the unknown environment through iterative interactions, or the offline setting where a policy is learned solely from logged data. However, each of these paradigms has inherent limitations: online algorithms suffer from high interaction costs and slow adaptation, while offline methods are constrained by dataset quality and lack of exploration capabilities. To address these complementary weaknesses, we propose hybrid CMAB-T, a new framework that integrates offline data with online interaction in a principled manner. Our proposed hybrid CUCB algorithm leverages offline data to guide exploration and accelerate convergence, while strategically incorporating online interactions to mitigate the insufficient coverage or distributional bias of the offline dataset. We provide theoretical guarantees on the algorithm’s regret, demonstrating that hybrid CUCB significantly outperforms purely online approaches when high-quality offline data is available, and effectively corrects the bias inherent in offline-only methods when the data is limited or misaligned. Empirical results further demonstrate the consistent advantage of our algorithm.
URL: https://openreview.net/forum?id=ZRzB7EVlui
---
Title: The Expanded Othello AI Arena: Evaluating Intelligent Systems Through Constrained Adaptation to Unseen Conditions
Authors: Byunghwa Yoo, Sundong Kim, KyungJoong Kim
Abstract: The ability to rapidly adapt to environmental changes is a core requirement for Artificial General Intelligence (AGI), yet most AI benchmarks evaluate performance in static environments. We present the Expanded Othello AI Arena, a benchmark designed to measure Skill-Acquisition Efficiency: the rate at which agents discover latent objectives and converge to effective strategies within a limited interaction budget. The Arena formalizes a spectrum of 56 environments using a parametric framework $\mathcal{E} = (\mathcal{L}, \mathcal{C})$, where $\mathcal{L}$ defines Othello board geometries and $\mathcal{C}$ represents latent winning conditions via a disc-ratio threshold $K$. This parameterization requires agents to decipher terminal rules through direct interaction while adapting against an opponent in a zero-sum setting; in narrow regimes, agents must precisely control their terminal occupancy under the hidden threshold, while terminal outcomes in which neither player satisfies the admissible interval are treated as draws. Unlike traditional evaluation, the Arena imposes a strict interaction budget to prioritize sample efficiency over asymptotic optimization. We establish the benchmark's utility through a neuroevolutionary adaptive-Minimax baseline that utilizes meta-learned spatial priors and adaptive weighting. Our empirical analysis reveals that while this baseline achieves competitive performance in standard and inverse regimes, it struggles in narrow-interval regimes that demand precise terminal control under latent objectives. Released as an extensible Python-based research toolkit, the Arena provides a standardized platform for exploring research directions, including test-time learning, in-context learning, and world models. Our project page is available at: https://expanded-othello.vercel.app
URL: https://openreview.net/forum?id=WXKQtqPC2d
---
Title: Learning with Differentially Private Sliced Wasserstein Gradients
Authors: David Rodríguez-Vítores, Clément Lalanne, Jean-Michel Loubes
Abstract: In this work, we introduce a novel framework for privately optimizing objectives that depend on sliced Wasserstein distances between data-dependent empirical measures. Our main theoretical contribution is a non-trivial analysis of the sensitivity of the Wasserstein gradients to individual data points, derived from an explicit formulation of the gradient in a fully discrete setting. This enables strong privacy guarantees with minimal utility loss. We demonstrate that standard privacy accounting methods naturally extend to Wasserstein-based objectives, allowing for large-scale private training. This supports a wide range of private machine learning applications involving distribution matching under privacy constraints on the source, the target, or both. These include: (i) an in-processing method for fairness mitigation using a private Wasserstein penalty, and (ii) what we believe is the first approach for training private sliced Wasserstein autoencoders. We validate our framework through experiments showing its ability to effectively balance privacy and utility, offering a theoretically grounded approach to privacy-preserving machine learning with sliced Wasserstein losses.
URL: https://openreview.net/forum?id=pqo5VNImZE
---
Title: Is Oracle Pruning the True Oracle? -- A Sanity-Check of Neural Network Pruning with Retraining
Authors: Sicheng Feng, Keda TAO, Huan Wang
Abstract: Oracle pruning, which selects unimportant weights by minimizing the pruned train loss, has served as the foundation for most neural network pruning methods for over thirty-five years, while few (if any) have thought about how much the foundation really holds. This paper, for the first time, attempts to systematically examine its validity on deep neural networks through empirical correlation analyses and provides meta-framework reflections on the field of neural network pruning. Specifically, this paper focuses on the pruning algorithms with three stages: training, pruning, and retraining. We analyze the correlation in model performance before and after the retraining stage. Extensive experiments (37K models are trained) across a wide spectrum of models (LeNet5, VGG, ResNets, ViT, MLLM) and datasets (MNIST, CIFAR10/CIFAR100, ImageNet-1K, MLLM data) are conducted. For large-scale experiments, we adopt approximate oracle pruning due to the prohibitive cost of exact oracle pruning. The results point to a counterintuitive conclusion: for deep learning models of nontrivial size (already at the scale of ResNet56 on CIFAR-10), pre-retraining performance is negligibly correlated with post-retraining performance. In other words, the weights identified by oracle pruning can scarcely guarantee strong performance following retraining. This further suggests that existing works that derive pruning criteria from oracle pruning may rest on a questionable foundational premise. Further studies suggest that rising task complexity is a primary factor behind the invalidity of oracle pruning nowadays. Finally, given the evidence, we argue that the retraining stage in a pruning algorithm should be accounted for when developing pruning criteria.
URL: https://openreview.net/forum?id=k13YnckIZY
---
Title: FedProTIP: Task-Agnostic Federated Continual Learning via Replay-Free Gradient Projection
Authors: Seohyeon Cha, Huancheng Chen, Haris Vikalo
Abstract: Federated continual learning (FCL) enables collaborative model training across distributed clients on sequentially arriving tasks without revisiting past data. However, existing approaches often suffer from catastrophic forgetting, rely on replay buffers or generative models that may violate privacy constraints, or assume knowledge of task identities during inference. We propose FedProTIP (Federated Projection-based Continual Learning with Task Identity Prediction), a replay-free FCL framework that maintains shared task-specific feature subspaces across clients. Each client extracts low-rank core bases from intermediate activations using randomized singular value decomposition, capturing dominant feature directions associated with the current task. These bases are transmitted to the server and aggregated to construct global task subspaces that capture shared feature directions across clients without requiring data sharing. During training, client updates are projected onto the orthogonal complement of previously learned subspaces to reduce cross-task interference and mitigate catastrophic forgetting. The learned subspaces are also reused during inference to estimate task identity via subspace relevance, enabling task-agnostic prediction without requiring explicit task labels. Experiments on CIFAR100, ImageNet-R, and DomainNet demonstrate that FedProTIP consistently outperforms state-of-the-art federated continual learning baselines while maintaining lower training time, memory footprint, and communication cost.
URL: https://openreview.net/forum?id=GW4aw0fUKC
---
New submissions
===============
Title: Do MLPs Inherit GNN Knowledge Uniformly? Class-Conditional Transfer Equity in Graph-Free Distillation
Abstract: Graph Neural Networks (GNNs) achieve strong performance by leveraging relational structure, but their reliance on message passing limits deployment in latency-sensitive settings. Distilling GNNs into Multilayer Perceptrons (MLPs) provides a feature-only student that approximates a graph-aware teacher without message passing at inference. Existing studies typically evaluate such transfer using average accuracy, implicitly assuming that teacher knowledge is inherited uniformly across labels. However, an MLP may match the teacher well on average while failing to inherit confident teacher knowledge for particular classes. We study this failure mode as class-conditional missed transfer in graph-free distillation. We first introduce \textit{Valid Transfer Agreement}, a label-level metric that measures whether a graph-free student preserves teacher decisions on test samples where the teacher is correct. To mitigate this failure mode, we propose \texttt{TED} (Transfer Equity Distillation), a training-time objective augmentation for existing GNN-to-MLP distillation methods. \texttt{TED} formulates reliability-guided distillation by treating correct and confident teacher decisions as transferable knowledge and defining label-wise transfer hardness as the portion of this knowledge not inherited by the student. By optimizing a differentiable hardness-based objective, \texttt{TED} reduces excessive missed transfer for difficult labels while leaving the teacher, student architecture, and inference procedure unchanged. Across eight homophilic and heterophilic graph benchmarks, \texttt{TED} reduces class-level transfer disparity and improves worst-label teacher--student agreement while preserving competitive accuracy and graph-free inference efficiency.
URL: https://openreview.net/forum?id=GDOC2rR4wv
---
Title: Mixed neural posterior estimation for simulators with discrete and continuous parameters
Abstract: Neural Posterior Estimation (NPE) enables rapid parameter inference for complex simulators with intractable likelihoods. NPE trains an inference network to estimate a probability density over parameters given data, typically assumed to be \emph{continuous}. However, many scientific models involve parameter spaces that are \emph{mixed}, that is, they contain both discrete and continuous dimensions. We address this limitation by extending NPE to mixed parameter spaces through an inference network that jointly handles discrete and continuous parameters. The inference network factorizes the joint posterior into discrete and continuous components, combining an autoregressive classifier for the discrete parameters with a generative model for the continuous parameters, trained jointly under a single simulation-based objective. In addition, we propose a diagnostic tool to assess the calibration of the mixed posterior approximation. Across tractable toy examples and real-world scientific simulators, our joint inference approach yields accurate and calibrated posteriors.
URL: https://openreview.net/forum?id=fOZZeRfdrC
---
Title: LLM-Guided Evolution of Symbolic Acceleration Inequality Families for MILP
Abstract: Integer programming (IP) is central to many combinatorial optimization tasks but remains challenging due to its NP-hard nature. A practical way to improve IP solvers is to manually design inequality families that exploit problem structure. However, this creative process requires deep expertise and has therefore been difficult to automate. Our proposed framework, Symbolic Acceleration via Generative Evolution (SAGE), automates symbolic discovery of acceleration inequality families at the modeling level. It reasons over a symbolic IP formulation and a natural language description of the problem to discover a reusable set of acceleration inequality families that can be instantiated for each problem instance. These acceleration inequality families are selected through empirical admissibility and impact screening rather than mathematical proofs. SAGE (i) initializes a population of candidate acceleration inequality families via an initializer agent that uses an LLM; (ii) it screens candidates for admissibility on a small development admissibility set for solution feasibility and the separation of the relaxation solution; and (iii) it iteratively refines the population of inequalities through evolutionary crossover and mutation agents. Across seven MILP benchmarks, SAGE reduces optimality gaps by up to 62.1% relative to baseline MILP formulations solved with a fixed time budget and reaches target gaps up to 5.51 times faster according to shifted geometric mean speedup. Comprehensive ablations show its robustness across different LLM backends and across solvers and internal cut settings.
URL: https://openreview.net/forum?id=rxJvINT074
---
Title: Mechanistic Interpretability of Transformer MLPs Through Exact Soft‑Gate Decomposition
Abstract: Transformer feed-forward (MLP) blocks perform nonlinear transformations, but their internal structure is often difficult to interpret using current methods. This paper presents an exact soft-gate decomposition that extracts the effective weight matrix for each MLP block at varying inputs. This approach generalises exact piecewise-affine decomposition from ReLU networks to Gaussian Error Linear Unit (GELU) and Sigmoid Linear Unit (SiLU) activations, covering both standard and gated MLP architectures. Verification across five models, including BERT, mBART, and GPT-2 for language tasks, ViT for vision tasks, and the gated SiLU architecture TinyLlama, confirms the accuracy of this method at floating-point arithmetic precision. The application of singular value decomposition to the effective weight matrix yields six structural metrics that characterise each block based on the dimensionality, intensity, and spectral concentration of its transformation. When applied to BERT, mBART, and GPT-2, and alongside ablation studies that provide causal ground truth, this analysis indicates that MLP computation is organised differently across encoder-only, encoder-decoder, and decoder-only models. Specifically, BERT blocks vary by transformation dimensionality, mBART blocks by transformation intensity, and GPT-2 blocks by intensity, which shows detrimental excess in the later layers. These structural differences, revealed only through the exact effective weight matrix, explain why scalar importance measures often fail to generalise across different architectures.
URL: https://openreview.net/forum?id=20lTANRDAz
---
Title: Sample-Based Constrained Inference for Matrix-Free Quantum Process Tomography
Abstract: Quantum process tomography reconstructs an unknown quantum channel from finite measurement counts. For sample-based uncertainty reporting, the sampled candidate channels should also remain physically valid, meaning completely positive and trace preserving (CPTP). Motivated by a flow-matching inverse-problem view, we introduce a matrix-free CPTP channel-sampling protocol for simulated full-basis process tomography. The protocol stores repeated tomography basis objects by index, constructs candidate channels with normalized Kraus factors, and calibrates scalar intervals using records from systems with one or two qubits before applying the fixed rule to four-qubit records. On a four-qubit controlled-NOT-family process, the center reconstruction reaches process infidelity 0.0143, average gate fidelity 0.9865, and passes projected CPTP diagnostics. Raw four-qubit intervals are under-dispersed, covering 0 out of 12 held-out scalar records, while the lower-qubit-selected calibrated intervals cover 12 out of 12 records at nominal 90\% level with mean width 0.0338. The calibrated four-qubit run stays within a declared 1800 second and 8192 MB resource envelope, taking 1287.49 seconds and 4000.04 MB. These results show physically valid four-qubit reconstruction with lower-qubit-calibrated scalar intervals under fixed acquisition in a controlled simulator setting.
URL: https://openreview.net/forum?id=f2LcKcJL2m
---
Title: Mixture-of-Gaussians-Guided Schedule Design for Brownian Bridge Diffusion Models
Abstract: Brownian Bridge Diffusion Models (BBDM) offer an appealing framework for image restoration and inverse problems by constructing a stochastic bridge from the clean signal directly to the degraded observation, rather than to pure noise. Despite their promise, the choice of bridge schedule is typically inherited from heuristics, and a principled analytical framework for schedule design has been lacking. In this work, we develop such a framework by offering a novel analysis of BBDM reverse dynamics under a Mixture-of-Gaussians (MoG) prior. This setting yields a closed-form ideal posterior and a corresponding MMSE denoiser, while the BBDM-induced reconstruction law is captured analytically through a tractable surrogate. Building on these expressions, we formulate two complementary schedule-design objectives: A Wasserstein criterion targeting perceptual quality and an MSE criterion targeting reconstruction fidelity. Our work exposes an inherent tradeoff between the two and proves the existence of universal schedules for both that are independent of the degradation and prior. Extensive experiments on controlled MoG settings confirm full alignment between theory and practice, and experiments on the FFHQ dataset across inpainting, deblurring, and super-resolution tasks validate the practical value of our schedule-design criteria.
URL: https://openreview.net/forum?id=NHccmvrBPO
---
Title: The Argmax Gap in Human Chess Move Prediction
Abstract: Human-move predictors are typically evaluated by whether their highest-ranked move matches the move played. This evaluation, however, does not distinguish between failing to identify the human move and identifying it as plausible but ranking another move first. We study this distinction on 884{,}049 positions from the Allie Lichess 2022 blitz test set using two strong, independently developed population-level human-move predictors, MAIA3 and Allie policy-only, evaluated under the same legal-move protocol. MAIA3 and Allie policy-only rank the observed human move in Top5 on 91.843\% and 90.932\% of positions, respectively, while reaching only 57.255\% and 55.734\% Top1 accuracy. We use the term \emph{argmax gap} for this discrepancy.
A diagnostic MAIA3--Allie oracle reaches 61.796\% Top1, but fixed selectors, pre-move learned selectors, and probability ensembles recover little of this headroom because their rescues are largely offset by newly introduced errors. The discrepancy persists across chess-context and model-uncertainty strata, while performance varies substantially across observed move-time and clock-context groups. Calibration, refinement, and ensembling improve NLL with little or no corresponding improvement in Top1. Because each position contributes only one observed move, these results do not establish how much of the discrepancy is reducible. They nevertheless show that ranking the observed move highly, fitting its probability well, and matching it exactly at Top1 are distinct evaluation outcomes.
URL: https://openreview.net/forum?id=AKQ2b7VLA4
---
Title: Scaling Online Complex Event Detection with Synthetic Supervision and Mamba-Based Neural Algorithmic Reasoning
Abstract: Modern machine learning models excel at detecting individual actions, sounds, or scene attributes from short, localized observations. However, many real-world tasks, such as in smart cities and healthcare, require reasoning over high-level \emph{complex events} (\emph{CE}s): spatiotemporal, rule-governed patterns of short-term \emph{atomic events}~(\emph{AE}s). Complex event detection (CED) is challenging due to long temporal dependencies, generalization beyond the training horizon, sparse \emph{CE}-level supervision, and cognitively demanding annotation, as \emph{CE} labels often depend on ordering, duration, negation, and completion-time semantics. These challenges are further amplified in an online setting that requires causal, streaming inference with limited computation.
We identify the primary bottleneck in online CED as learning robust \emph{CE} rules, and propose a Neural Algorithmic Reasoning framework that decouples rule learning from low-level sensor semantics by (i) generating large-scale synthetic \emph{AE}-level concept traces to pretrain a Mamba-based \emph{CE}-rule reasoner, and (ii) introducing an adapter that learns to map raw sensor inputs into the reasoner’s latent space using limited, labeled sensor data. We introduce a controlled online multilabel CED testbed built from real-world multimodal sensor clips and rule-generated \emph{CE} labels, with stress-test settings that vary sensor noise, distribution shift, and the window size used to segment streaming sensor sequences. Experiments show that our method matches or exceeds the strongest baselines under these stress tests and longer-horizon generalization, with $5\times$ fewer labeled sensor sequences and $10$--$20\times$ fewer FLOPs. Code and dataset available at: \href{https://anonymous.4open.science/r/naroce_dailyoce-C7F6/}{/r/naroce-C7F6/}.
URL: https://openreview.net/forum?id=UJwnFHlqEf
---
Title: A Journey Through Surface Convolutions
Abstract: Surface data appears in a wide range of machine learning tasks. Among other objectives, Geometric deep learning generalizes Euclidean CNNs to two-dimensional Riemannian manifolds, allowing to learn patterns of data along curved surfaces. While the literature presents numerous surface convolutions, existing comparisons often vary both the convolution and the underlying surface charting algorithm, complicating the isolation of the source of performance differences. Furthermore, unlike in Euclidean CNNs, the size of the convolution neighborhood is not implied by the template size but manually set in the form of a chart radius. Finding appropriate chart radii has received little attention and thus remains poorly understood. To address these challenges, we (i) provide a structured survey of surface convolutions and (ii) reformulate several approaches into a unified discretization notation. Using this notation, we (iii) develop a charting-method agnostic surface CNN framework, allowing to isolate downstream effects of charting algorithms from those of the convolution itself, and to systematically study the impact of chart radii. We provide a Python library and (iv) use it to conduct an ablation study encompassing 420 surface CNN configurations. We find that varying the chart radius gives rise to statistically significant performance changes in nearly all of our experiments. Additionally, we find statistically significant effects of charting algorithms in local prediction tasks. These findings suggest that both the choice of the charting algorithm and chart radius constitute important experimental factors. Future comparisons among surface CNNs should thus control charting effects by using a common charting algorithm and treat chart radii as tunable hyperparameters.
URL: https://openreview.net/forum?id=lCwv0bo973
---
Title: Impact of Fairness Regulations on Institutions' Policies and Population Qualifications
Abstract: In this paper, we consider institutions whose primary objective is to maximize utility by selecting the most qualified individuals. We propose a regulatory framework to promote demographic parity by penalizing the disparity in the institution's selection across groups.
We analyze how the shape of the penalty function and specific conditions determine the effectiveness of a discrimination penalty in reducing selection disparities. Additionally, we explore the implications of such a penalty when individual qualifications evolve over time in response to the penalization policy. We identify scenarios where the penalty could hinder the natural attainment of equity within the population. Moreover, we identify conditions that counteract this undesirable outcome and ensure fairness.
URL: https://openreview.net/forum?id=oXZOIFH5te
---
Title: Automated Synthesis of RL Environments and Reward functions from Policy Narratives via Grammar-Constrained LLM Refinement
Abstract: Translating policy documents into executable reinforcement learning environments requires grounding normative language into typed state spaces, action sets, and reward functions — a step that currently demands weeks of expert engineering and produces rewards reflecting the engineer’s interpretation rather than the text. NarraRL addresses this by applying grammar-constrained large language model (LLM) extraction to map policy clauses to a typed, clause-traceable RLSpec schema, then synthesizing Gymnasium environments with co- generated reward functions. Evaluated on all 17 UN Sustainable Development Goal domains with three small open-weight LLMs (1.5B–3.8B), four findings emerge. Grammar-constrained extraction achieves significantly higher Clause-Reward Overlap (CRO) than unconstrained prompting ($\Delta= + 0.052$, Wilcoxon $p=1.4×10−26$, $n=153$; $k$-stable: $\Delta \in [0.052, 0.053]$ for $k \in \{1, 3, 5, 6\}$). Synthesis quality depends critically on template scaffolding: across all 17 SDGs (153 conditions), NarraRL synthesizes executable environments at a 99% pass rate with near-zero repair retries, yet standardized multi-LLM evaluation ($n=306$) exposes near-zero Gymnasium compliance -Stage-1 syntax failures account for 78% of breakdowns- and a controlled ablation confirms the causal mechanism: template scaffold yields 100% vs. 2.4% for LLM-direct codegen under identical conditions (Fisher $p=7.6×10^{−94}, n=340$). Co-generated rewards match hand-designed rewards in mean policy performance (Mann-Whitney $p=0.807$, n=20; Levene $p=0.414$), while semantically misaligned rewards produce 5× higher variance (σ=3.36, non-overlapping 95% confidence interval, CI). Finally, on an SDG 3 policy-discovery case study, the naive near-identity dynamics give a deep Q-network (DQN) no advantage over random $(p=0.849, n=30$); wiring the LLM’s extracted action intent into the environment makes it solvable, and adding a single domain-agnostic opportunity-cost term makes it learnable — DQN then satisfies ≥ 3/7 health targets in 100% of seeds and outperforms random ($p=5.8×10^{−7}$), matching a hand-designed expert matrix ($p=5.3×10^{−7}$) with no human in the loop. This isolates environment dynamics —not reward extraction— as the decisive factor, and shows it can be supplied automatically. These results establish grammar-constrained extraction and template scaffolding as reproducible, clause-traceable primitives for converting policy text into RL environments.
URL: https://openreview.net/forum?id=YikG1amX3O
---
Title: Finite-sample Calibration Guarantees for Binary Segmentation
Abstract: Machine learning–based segmentation models are frequently miscalibrated, producing for each pixel a label prediction whose associated confidence does not accurately reflect the true underlying probability. This limitation is particularly critical in high-stakes applications such as medical image segmentation. Although several calibration methods have been proposed, only a small subset of them provides theoretical guarantees on the calibration error, i.e., they do not typically ensure that, with high probability, the model exhibits a low calibration error. In this work, we develop a post-hoc calibration method for which formal calibration guarantees can be established. To achieve this, we derive from standard concentration inequalities an improved theoretical bound that explicitly accounts for the spatial dependence inherent in binary segmentation outputs. We empirically demonstrate that the proposed calibrator satisfies the derived guarantee and outperforms existing post-hoc calibration approaches.
URL: https://openreview.net/forum?id=fHHbzZZ2Na
---
Title: Structuring Relations Among Learning Paradigms via Protocol–Objective–Resource Reductions
Abstract: Modern machine learning increasingly spans supervised, transfer, continual, meta-learning, and other training regimes, often while reusing the same hypothesis families, architectures, and optimizers. Yet these regimes can differ substantially in information access, success criteria, memory assumptions, adaptation protocols, and sample accounting. As a result, it is often unclear whether a proposed paradigm is genuinely distinct, a special case of an existing paradigm, or part of a broader structural relationship among paradigms. This motivates a structural theory for comparing learning paradigms while separating representational capacity from protocol, objective, and resource choices. We develop a protocol--objective--resource (POR) framework for making such comparisons precise. In this framework, a learning paradigm is specified by its environment class, observation protocol, admissible learner class, performance functional, and resource accounting rule. We define POR reductions through environment embeddings, learner compilers, threshold maps, and calibrated sample-budget overheads. Our main theorem shows that a POR reduction implies worst-case sample-complexity domination on embedded comparison classes, transferring upper bounds in one direction and lower bounds in the other. We further provide reusable templates for proving reductions, show why explicit accuracy regimes are necessary to avoid vacuous comparisons, and prove that strengthening the objective can strictly increase minimax sample complexity even when the protocol and learner class remain unchanged. We instantiate the framework on supervised, transfer, continual, and meta-learning abstractions. Under explicit closure assumptions, the framework yields exact reductions showing that continual learning contains transfer learning as a special case, transfer learning contains supervised learning as a special case, and meta-learning also contains supervised learning as a special case. The same formalism captures monotone refinements within a paradigm, for example, as replay memory or task identifiers vary in continual learning. Together, these results provide a principled structure for relating learning paradigms and enable the formal transfer of understanding and methods across them.
URL: https://openreview.net/forum?id=8ZyXUFRfb2
---
Title: Reinforcement Learning for Symbolic Equation Solving
Abstract: We present, to our knowledge, the first reinforcement-learning agent that solves \emph{nonlinear} and \emph{open} symbolic equations step-by-step from reward alone -- with no supervised solution traces. Prior RL for equation solving is limited to linear equations~\citep{poesia2021contrastive,dabelow2024symbolic}, while one-shot neural solvers~\citep{lample2020deep} require ${\sim}10^8$ supervised pairs and emit an answer rather than a verifiable procedure. We cast algebra as an MDP with a \emph{dynamic, per-state action space}, solved by a tree-structured policy network (\textbf{TreeMLP}), and reach \emph{open} equations -- those requiring a change of variables (CoV), e.g.\ completing the square -- by exposing CoV as a macroaction whose \emph{trigger} the agent learns; the substitution itself comes from a supervised generator interchangeable with a computer-algebra call (${\le}0.02$ end-task effect), so the learned component is the trigger, not the substitution. The agent solves quadratic, cubic, quartic, and exponential open families ($5$-seed beam-decoded test accuracy $\mathbf{0.79}$ vs.\ $0.31$ baseline; greedy ${\approx}0.24$, so the result is decode-time search informed by the policy) and matches the prior best on closed CommonCore equations ($0.93$ vs.\ ConPoLe's $0.925$). At $10\times$ scale we additionally characterize -- and partially mitigate with a learning-progress curriculum -- a seed-level bimodality in which an otherwise-identical run either learns the task or collapses into a class-level policy trap.
URL: https://openreview.net/forum?id=JlC5BJiikD
---
Title: Markovian ODE-guided scoring can assess the quality of offline reasoning traces in language models
Abstract: Reasoning traces produced by generative language models are increasingly used for tasks ranging from mathematical problem solving to automated fact checking. However, existing reasoning evaluation metrics are typically validated using either synthetic perturbations or benchmark-specific correlations. Moreover, the design choices underlying many metrics are often closely tied to the perturbation sets used for their evaluation, raising concerns about their generalizability. Additionally, despite employing multiple perturbation types, prior work frequently relies on binary grading schemes that provide only a coarse assessment of reasoning quality. Consequently, such evaluations offer limited evidence that a metric remains sensitive to realistic degradations in reasoning quality or that it generalizes across reasoning domains. As a result, it remains unclear whether existing metrics capture properties that align with human judgments of reasoning quality. To this end, we introduce MarODE, an offline evaluation framework that assigns quality scores to reasoning traces. Its effectiveness is assessed using human-centric perturbations and human judgments, which jointly evaluate the fundamental dimensions of an evaluation metric – goodness and soundness. The approach is grounded in a Markovian formulation of reasoning progression and an ordinary differential equation based characterization of trace dynamics, enabling efficient evaluation of reasoning quality. In a large-scale evaluation, MarODE outperforms existing baselines consistently under Somers’ D correlation. Our results emphasize the value of theory-driven evaluation frameworks as reasoning traces become central to language model-based systems.
URL: https://openreview.net/forum?id=RfAEAWk48M
---
Title: Residual Clock Accumulation and Rotary Decoherence under Time Warping
Abstract: Rotary sequence models rely on an internal notion of position that should remain coherent with the sequence parameter being modeled. This paper studies a failure mode that arises when observed index time is a warped version of the latent sequence clock: a fixed or learned clock can accumulate residual phase error, causing rotary frequency bands to decohere over long offsets. We formulate this phenomenon as a residual-clock accumulation problem. If a model estimates the true local warp speed \(c'(t)\) with residual \(\eps(t)=\widehat c'(t)-c'(t)\), then the quantity governing rotary band error at offset \(u\) is the accumulated residual \(W_u=\sum_{r=t}^{t+u-1}\eps(r)\). We derive a two-moment prediction of the decoherence curve using empirical \(\mu(u)=\E[W_u]\) and \(v(u)=\Var(W_u)\), and use the empirical characteristic function of \(W_u\) only as a diagnostic for moment-closure failure. Using controlled synthetic warp generators with known ground-truth clock speed, we evaluate fixed clocks, static-oracle clocks, MLP clocks, Mamba-\(\Delta\)-based clocks, and frozen-\(\Delta\) controls. The two-moment prediction matches observed decoherence horizons with \(R^2\) near \(0.9\) and median relative error around \(8\%\) in both weakly supervised and forecast-only settings. Its failures occur in structured residual regimes where higher-order or non-Gaussian window-sum behavior breaks the moment closure. Under increasing warp-distribution shift, fixed and frozen clocks are warp-limited from the start, while learned adaptive clocks can delay the transition from discrimination-limited to warp-coherence-limited behavior. These results identify residual-clock accumulation as a measurable mechanism behind rotary decoherence under time warping.
URL: https://openreview.net/forum?id=XWc1HxVXNE
---
Title: LLM-Driven Algorithm Design for Quantum Circuit Synthesis based on Binary Decision Diagrams
Abstract: Quantum circuits are central to implementing quantum algorithms on quantum devices, where quantum gates must be reversible. Many quantum algorithms rely on Boolean functions, which must therefore be implemented reversibly within quantum circuits. Reversible circuit synthesis provides a way to translate such Boolean functions into reversible circuits. Binary decision diagrams (BDDs) offer a scalable approach to this task, but the resulting BDDs and circuits depend heavily on variable ordering. Existing ordering heuristics commonly minimize BDD size because it is closely tied to the circuit size. However, BDD size is an imperfect proxy for the quantum cost of the synthesized circuit (QCC). We propose QuantumEvo, an evolutionary framework that uses an LLM as a heuristic generator for QCC-aware BDD variable ordering. Instead of predicting orderings directly, QuantumEvo searches over ordering heuristics initialized from multiple heuristic families. Candidate heuristics directly manipulate variable orderings using standard BDD operations and are selected by downstream QCC. The discovered heuristic, HGA-QE, modifies the sifting step inside a genetic algorithm so that the procedure is better aligned with QCC. Across the benchmark set, HGA-QE is strictly best on 13.5% of functions and achieves a 70.9% tie-or-win rate against the per-function best baseline.
URL: https://openreview.net/forum?id=iL7YeQp7je
---
Title: Towards Principled Test-Time Adaptation for Time Series Forecasting
Abstract: Test-time adaptation (TTA) has recently emerged as a promising approach for improving time series forecasting (TSF) under distribution shift. Existing TSF-TTA methods utilize revealed targets in different ways, leading to heterogeneous adaptation protocols that lack a clearly unified formulation. To address this issue, we revisit TSF-TTA through the lens of protocol cleanliness and introduce a principled adaptation protocol based solely on matured ground truth. Under this protocol, we analyze existing adapters in the frequency domain and reveal that their prediction corrections typically exhibit limited and weakly structured spectral modifications. Driven by this observation, we propose Frequency-Aware Calibration (FAC), a lightweight calibration method that directly parameterizes prediction corrections in the frequency domain. Across diverse datasets, forecasting horizons, and source forecasters, FAC achieves competitive and consistent performance while requiring substantially fewer trainable parameters than the compared TSF-TTA adapters.
URL: https://openreview.net/forum?id=bA2bMEOUFL
---
Title: Exploring Concept Subspace for Self-explainable Text-Attributed Graph Learning
Abstract: We introduce Graph Concept Bottleneck (GCB) as a new paradigm for self-explainable text-attributed graph learning. GCB maps graphs into a subspace---a \textit{concept bottleneck}---where each concept is a meaningful phrase, and predictions are made based on the activation of these concepts. Unlike existing interpretable graph learning methods that primarily rely on subgraphs as explanations, the concept bottleneck provides a new, semantically grounded form of interpretation that is human-readable. To construct this concept space, we contrastively pretrain a graph encoder so that graphs are aligned with a concept embedding space, employ large language models to propose candidate concepts at both the instance and the class level, and apply the information bottleneck principle to retain a compact, causally informative subset for the downstream predictor. This pipeline not only yields more concise and faithful explanations but also explicitly guides the model to ``think'' toward the correct decision. We empirically validate GCB on five text-attributed node classification benchmarks, and find that GCB achieves intrinsic interpretability with accuracy on par with black-box Graph Neural Networks (GNNs). Moreover, GCB's advantage over GNNs increases under distribution shifts and adversarial perturbations, suggesting that its robustness arises from concept-guided prediction rather than purely graph-structural reasoning.
URL: https://openreview.net/forum?id=3QS5avNVHn
---
Title: Stay Rotated: Training-Free Universal Codebook Quantization for Large Language Models via Orthogonal Invariance
Abstract: We present a training-free quantization method for large language model (LLM) inference that exploits the orthogonal invariance of inner products to derive a universal codebook applicable to any pretrained model without calibration. After applying a randomized Hadamard rotation, all weight coordinates converge to N(0, 1/d) — a result we quantify with a Berry-Esseen bound giving O(1/√d) convergence rate. The Lloyd-Max optimal quantizer for this universal distribution yields a 30-byte fixed codebook (b=3 bits, 8 centroids) that is independent of model, layer, or weight type. We further observe that keeping computations in the rotated domain — never explicitly inverting the rotation — eliminates the dequantization overhead that rotation-based methods such as QuaRot and SpinQuant incur, by exploiting orthogonal invariance ⟨x, w⟩ = ⟨Rx, Rw⟩. This enables matrix-vector products to be computed via table lookups and additions rather than scalar multiplications. We prove an error bound via Cauchy-Schwarz that matches empirical SQNR within 0.1 dB across all 224 weight matrices of Llama-3.1-8B. Across four model families spanning 3B to 109B parameters — Llama-3.2-3B, Llama-3.1-8B, Qwen3-8B (a non-Llama architecture), and the 109B/17B-active MoE Llama-4-Scout (768 experts) — the same 30-byte codebook achieves competitive perplexity (4-bit ΔPPL = +0.51 on Llama-3.1-8B WikiText-2) and downstream accuracy without any calibration data. Open-source implementations on NVIDIA H100 GPU and AMD EPYC 9654 CPU validate that the algorithm is practical on diverse hardware substrates.
URL: https://openreview.net/forum?id=wwNZeCMTxL
---
Title: Unlocking Out-of-Distribution Generalization in Transformers via Latent Space Reasoning
Abstract: Systematic, compositional generalization beyond the training distribution remains a core challenge in machine learning—and a critical bottleneck for the emergent reasoning abilities of modern language models.
This work investigates out-of-distribution (OOD) generalization in Transformer networks using a GSM8K-style modular arithmetic circuit-evaluation task as a testbed.
We introduce and explore a set of four architectural mechanisms aimed at enhancing OOD algorithmic generalization: (i) *input-adaptive recurrence*; (ii) *algorithmic supervision*; (iii) *anchored latent representations via a discrete bottleneck*; and (iv) *an explicit error-correction mechanism*.
Collectively, these mechanisms yield an architectural approach for native and scalable latent space reasoning in Transformer networks with robust algorithmic generalization capabilities.
We complement these results with mechanistic interpretability analysis showing how these mechanisms yield robust OOD generalization.
URL: https://openreview.net/forum?id=8bFgEyRLrO
---
Title: CaliBench: Are the Stochastic Dynamics of Video World Models Physically Calibrated?
Abstract: Video world models are designed to approximate the stochastic distribution of physical outcomes through generative sampling, but existing benchmarks evaluate properties of each individual generation or aim to compare output distributions in a coarse grained way over an entire dataset, and do not investigate the fine grained aleatoric uncertainty of specific physical phenomena. We introduce CaliBench, a benchmark that tests this directly by scoring outcomes in a physically interpretable discrete space — a bin index, a die face, a suit, a colour — rather than in a learned feature space (e.g. that of FID), where we are able to directly calculate the distributional distance of the model outputs from a known reference distribution. This is possible because we carefully curate outcome spaces where the reference distribution is known in closed form (binomial Galton boards, Bernoulli forks, uniform dice/cards/lottery, a known-skewed European-roulette colour), which enables an exact calibration test which is not dependent on an empirical proxy. We decompose performance into two orthogonal axes that a single accuracy metric would conflate: scorability (the fraction of video generations producing a scoreable outcome) and calibration (the total variation distance from the reference distribution on the scoreable sample). We also consider the statistical significance of miscalibration using a $\chi^2$ test. We apply the protocol to nine scenes and six state-of-the-art image-to-video models (WAN-2.7, SeeDance-2.0, HappyHorse-1.0, Veo 3.1, Runway Gen-4.5, Cosmos3-Super) on $32$ video generations each. Results reveal a consistent pattern: models concentrate output probability mass on a small subset of outcomes rather than reproducing the reference distribution. Most scene–model combinations are significantly miscalibrated, in the most extreme case collapsing entirely to a single outcome, for example Veo 3.1 on dice. On roulette, generated videos often leave the ball ambiguously placed, so several models also have low scorability. Performance varies sharply by scene rather than by model: no single model dominates across all nine scenes. We release the protocol and a metric (mean normalised total variation, mnTV) to enable comparison of new models with the results in this paper.
URL: https://openreview.net/forum?id=rWE29rkvDz
---
Title: PairAlign: A Framework for Sequence Tokenization via Self-Alignment with Applications to Audio Tokenization
Abstract: Modern learning systems map raw sensory signals into continuous vectors, yet many operations---comparison, memory, retrieval, and reasoning---are naturally expressed over discrete symbolic structures. In language, this interface is given a priori through tokens; for continuous sensory signals such as speech and audio, it must be learned. Recent audio tokenizers often rely on geometric vector quantization, semantic clustering, or codec-style reconstruction, assigning tokens locally at a frame or short-window level. Consequently, sequence-level properties such as cross-realization consistency, compactness, learned length control, termination, and edit-based similarity are rarely optimized directly.
We introduce PairAlign, a framework for learning compact audio token sequences through sequence-level self-alignment. PairAlign treats tokenization as conditional sequence generation: an encoder maps speech to a continuous conditioning representation, and an autoregressive decoder generates the complete token sequence from BOS, learning token identities, ordering, length, and EOS placement. Given two content-preserving views of a segment, each view's token sequence is trained to receive high likelihood under the other's representation, while unrelated in-batch examples provide competing symbolic sequences. This gives a scalable surrogate for edit-distance preservation while discouraging many-to-one collapse.
PairAlign combines a staged transition from VQ-style geometric tokenization to adaptive EMA-teacher sequence tokenization with cross-paired teacher forcing, prefix corruption, encoder-summary conditioning, structured self-attention dropout, hardest-K likelihood contrast, repetition-aware target generation, length-constrained decoding, and post-hoc timing recovery from cross-attention.
PairAlign learns compact, non-degenerate token sequences with broad vocabulary usage and strong cross-view consistency. In retrieval, it operates at 12.71 tokens/s and reduces the archive token count by approximately 55% relative to the baseline VQ-style geometric tokenizers, while preserving meaningful edit-distance search. These results reveal a compactness--locality trade-off: PairAlign gives up dense frame-level redundancy and does not claim to dominate high-rate geometric or pretrained SSL-based tokenizers on every local retrieval metric, but obtains a lower-rate symbolic interface for comparison, retrieval, and structural analysis.
More broadly, PairAlign is an audio instantiation of sequence-symbolic predictive learning: like JEPA-style objectives, it predicts an abstract target associated with another view rather than reconstructing raw input, but the target is a learned variable-length symbolic sequence rather than a continuous latent. This points toward self-supervised symbolic-interface induction for continuous inputs requiring compact, stable, discriminative, and adaptively sized token sequences.
URL: https://openreview.net/forum?id=gd0IwqmO71
---
Title: Keep Your Boundaries: From Finite Elements to Simplicial Convolution
Abstract: Diffusion on graphs can be viewed from a perspective of partial differential equations (PDEs), while Graph Neural Networks (GNNs) can be interpreted as a discrete counterpart of diffusion PDEs. As such, many emerging GNNs have been inspired by PDE solvers from numerical analysis. However, most PDE-assisted GNNs employ finite-difference methods, which essentially consider only the function discretization, evaluated on the grid nodes. We propose to bring the ideas of finite element methods (FEMs) from numerical PDEs to GNNs, allowing us not only to define the function on the local region (simplices) but to enforce the appropriate {continuity} constraints on the neighboring simplices. We develop a novel Simplicial Element Network (SEN) framework with the dual simplicial convolution encoder, based on a {\it specially-designed interior boundary matrix} to enhance the graph representation learning. We discuss theoretical underpinnings of SEN, using the notion of combinatorial Laplacians on simplices and discrete Hodge theory. Our experiments indicate that SEN outperforms 20+ state-of-the-art baseline methods on 10+ benchmark datasets on 5 tasks: protein sequence recovery, link prediction, air quality forecasting, trajectory prediction, and fluid flow reconstruction.
URL: https://openreview.net/forum?id=NlA5gRc6f7
---
Title: Generative Retrieval via Diffusion Transformer with Metric-Ordered Sequence Training and Hybrid-Policy Preference Optimization
Abstract: Embedding-based retrieval ranks items by their similarity to a query in a shared vector space and usually aims to return the highest-scoring items. In many production settings this is not what is wanted: given a seed set that expresses a fine-grained pattern, one needs more items that both satisfy a target attribute and stay within that pattern. We formalize this as pattern-preserving attribute retrieval. The two goals pull against each other: averaging the seeds preserves the pattern but stays in a low-attribute region, while global attribute retrieval drifts to unrelated patterns. We approach the task with continuous generative retrieval, where a model reads a sequence of item embeddings and generates one or more query embeddings for nearest-neighbor search. We propose MO-DiT+HPPO (Metric-Ordered Diffusion Transformer with Hybrid-Policy Preference Optimization, HPPO), a staged framework with large-scale raw-sequence pretraining, multi-domain metric-ordered continuation pretraining, domain-specific tail-centroid supervised fine-tuning, and HPPO. The metric-ordered stages convert sparse online retrieval labels into in-pattern trajectories ordered from low to high predicted attribute density, so a single continuation-pretrained model learns the metric-improvement direction across all domains. HPPO then aligns the final query distribution with the true online objective: in each round it forms a hybrid-policy candidate pool — deterministic high-metric constructions together with the current policy's own guidance-scale fan — labels every candidate with the real online intersection metric Joint@K, and applies iterated, reference-anchored preference optimization with held-out-validation early stopping; a simple tradeoff-aware Pareto pair filter then keeps only winner pairs that do not lower same-pattern purity, so updates raise the attribute metric without sacrificing the pattern. Across four large-scale attribute domains under strict item- and pattern-holdout protocols, metric-ordered DiT training improves the primary intersection metric over a strong pretrained generative retriever, and HPPO improves it further, with gains that are large and paired-bootstrap significant on seven of the eight domain-split cells and a marginal tie on the hardest domain's pattern split. The Pareto filter is the key ingredient: on most domain-split cells it raises Joint@K while keeping the same-pattern share higher than an unconstrained preference policy, pushing the attribute-pattern frontier outward rather than merely sliding along it. Metric-predictor validation, matched order ablations, CPT/SFT comparisons, and a candidate-policy ablation show where the gains come from and that static and policy-generated candidates are complementary.
URL: https://openreview.net/forum?id=aUX1pxM5lF
---
Title: Verified Self-Improvement and Test-Time Scaling on One Consumer GPU: A Reproduction Study with Confirmations and a Substitution Effect
Abstract: Small, locally hosted language models remain unreliable on multi-step reasoning tasks. Most
reported improvements: verified self-training, test-time scaling (TTS), and confidence-based
routing, have been evaluated on full-precision models running on server-class hardware. We
revisit these claims under realistic consumer constraints: a single 4-bit 9B model running on
one consumer GPU (RTX 5080, 16 GB). We reproduce four published results from Think–
Prune–Train, self-consistency, s1 budget forcing, and on-device routing. We also perform
an additional comparison not studied in Think–Prune–Train: self-training versus frontier-
teacher distillation. All reported gains replicate on consumer hardware. We then report
two findings. N1: Fine-tuning and test-time scaling act as substitutes. A weak base model
(81.4% sampled maj@1; ≈80% greedy) and a self-trained model (86.4%) both converge
to around 94.0% under maj@16. Additional training improves low-compute performance,
but its advantage disappears as test-time compute increases. The same pattern holds on
progressively harder benchmarks, from GSM8K to MATH-500 and AIME. N2: a GSM-8K
comparison of self-training vs frontier-teacher distillation with cross-family GPT-5.5 (89.5%)
and same family Qwen3.5-27B (90.0%) and Qwen3.5-397B (92.5%) models. We release all
code used in the study in https://anonymous.4open.science/r/single-gpu-tts-tpt-0007.
URL: https://openreview.net/forum?id=iV2EnlDFwe
---
Title: Which Examples Preserve Decisions? Verified Inverse Design for Model Selection
Abstract: Model selection recurs across deployment slices and robustness tests, while training and evaluation data can be expensive to acquire, retain, or repeatedly process. We study decision-preserving acquisition (DPA): given a fixed model-selection protocol and an observed multi-scenario winner map, identify compact training data support that reproduces the same winners with prescribed margins after retraining. Unlike coreset or data-evaluation objectives, DPA targets the stability of the model-selection decision itself. In an idealized convex ERM regime, DPA admits a Karush--Kuhn--Tucker (KKT) feasibility view, while the associated cost-minimization problem is combinatorial; these structural facts motivate local proposal models combined with direct verification. We propose VERDICT (VERified Decision Inverse Consistency Training), a propose-and-verify method for DPA. VERDICT derives a local decision-response operator, uses trust-region surrogates over working sets to propose lower-cost designs, and accepts a candidate only after rerunning the fixed protocol verifies the target margins under the stated decision convention. We evaluate VERDICT in settings where verification is feasible, including decision-preserving subset selection under a train-and-verify setting and downstream subset selection for evaluation.
URL: https://openreview.net/forum?id=rWz4DKGsw9
---
Title: Data Selection for Reward Labeling in Limited-Feedback Reinforcement Learning
Abstract: The ability of reinforcement learning (RL) algorithms to learn effective policies is determined by the rewards available during training. However, for practical problems, obtaining large quantities of reward labels is often infeasible due to computational or financial constraints, for instance, when relying on human feedback. When reinforcement learning must proceed with limited feedback---only a fraction of samples get reward labels---a fundamental question arises: *which* samples should be labeled for RL training to maximize policy performance? We formalize this problem of *reward selection* for reinforcement learning from limited feedback, introducing a new problem formulation that facilitates the study of strategies for selecting impactful rewards. Various types of selection strategies are investigated: (i) ones that rely on reward-free information such as state visitation and partial value functions, (ii) ones that perform near-optimal selection requiring prohibitive costs, and (iii) ones that trade off those costs for some loss in performance. We find that critical subsets of rewards are those that (1) guide the agent along optimal trajectories, and (2) support recovery toward near-optimal behavior after deviations. Effective selection methods yield near-optimal policies with significantly fewer reward labels than full supervision, establishing reward selection as a powerful paradigm for scaling reinforcement learning in feedback-limited settings.
URL: https://openreview.net/forum?id=N8OlHGeHSy
---
Title: Transformer Depth Reduction via Low-Pass Filtering
Abstract: We introduce Time-Variable Low-Pass (TVLP) filtering of attention projections and apply it to a GPT-3 small-style Transformer. We implement TVLP with custom sequential Metal kernels on Apple M4 Pro (MPS) and demonstrate that a depth-11 model with TVLP applied across all layers exhibits lower bits-per-byte (BPB) than the depth-11 baseline and nearly matches the depth-12 baseline at equal training steps throughout the Chinchilla regime. Measured at equal FLOPs, the TVLP-enhanced depth-11 model shows lower BPB than both baselines, while at equal wall-clock time it achieves lower-or-equal BPB throughout the training process. We also implement TVLP with custom warp-scan CUDA kernels and confirm the wall-clock advantage at training speeds measured on Nvidia L40S. Notably, the parameter count, inference cache size, and FLOPs per token added by TVLP across all 11 layers represent only 2.42% of the number of parameters, 1.61% of the size of the KV-cache, and 2.11% of the FLOPs per token in a single standard Transformer block, respectively.
URL: https://openreview.net/forum?id=gkVqkeNA0V
---
Title: Languages Change, Safety Weakens: Jailbreaking LLMs Across English and Indian Languages
Abstract: Jailbreaking poses a critical safety risk in large language models (LLMs), yet its impact on low-resource multilingual settings remains underexplored. In this study, we examine multilingual jailbreaking vulnerabilities in LLMs across five languages: Hindi, Marathi, Telugu, Bengali, and English. We introduce Indic-JailbreakBench, a dataset comprising 1,668 malicious prompts per language across 12 categories, incorporating culturally grounded harmful prompts. Additionally, we propose four novel jailbreaking techniques: (a) history-based prompting, (b) biasing the attention of LLM towards unnecessary task, (c) Shakespearean-style prompt rewrites, and (d) contextual injection via harmful news articles. We employ a multi-LLM judging mechanism, along with Guard models and Keyword-ASR to assess attack success rates (ASRs) to evaluate eight state-of-the-art open-source LLMs. We also show that LLM responses are safer when prompted through its corresponding chat template compared to scenarios without it in almost all settings. Our findings reveal significant disparities in safety mechanisms across languages, with Indian languages often exhibiting higher susceptibility to jailbreak attempts than English. Additionally, three of our proposed jailbreaking strategies achieve very high attack success rates (ASR), exceeding 85% across all languages with highest ASR being 93% through Shakespearean prompt styling for English language.
URL: https://openreview.net/forum?id=TioFTA4y9E
---
Title: What Predicts Correctness in Text-to-SQL? A Selective-Prediction Study
Abstract: Evaluating uncertainty in AI-generated SQL queries requires an estimate of whether it is correct, where
correct means the query executes to the same result as a human-written reference query. We study
which signals provide correctness on hard multi-table text-to-SQL using \auroc{} to compare how well various methods rank correct queries above incorrect ones. Using BIRD and Spider SQL query datasets, the black-box statistical signals such as string, structural, and execution self-consistency, a schema-relevance score,
and query executability all fall between about $0.61$ and $0.68$ \auroc{}, with string
self-consistency the strongest at $0.675$. White-box log-probability is similar ($0.67$). The
signals that move past this ceiling are verification-based, with an LLM judge scoring from $0.72$
(GPT-4o-mini) to $0.78$ (Claude). Judges from different providers make different errors, so a two-provider ensemble reaches $0.82$ \auroc{} with a well-calibrated
probability (expected calibration error $0.03$), and it supports useful abstention frontiers (for
example, answering $27\%$ of questions at $24\%$ selective risk) where self-consistency offers no
valid low-risk subset. The pattern holds across two benchmarks (BIRD and Spider), two generators, and
two judge providers. We also explore whether a verifier can be trained. Fine-tuned verifiers, both encoder and generative, reach about
$0.77$--$0.79$ \auroc{} in-distribution but fall to about $0.66$ on unseen schemas, and scaling to
7B, adding schema diversity, distilling a strong judge's rationales, and cross-benchmark training all
fail to close that gap. Cross-schema transfer appears to track model scale and reasoning rather than
fine-tuning. In practice, correctness uncertainty for text-to-SQL lives in reasoning-based signals, meaning that a
fine-tuned verifier is a good in-domain tool, while a verifier that generalizes across schemas
currently means a large frozen reasoning model.
URL: https://openreview.net/forum?id=ogZu6QM6tY
---
Title: ESLM: Risk-Averse Selective Language Modeling with Hierarchical Batch Selection
Abstract: Large language model pretraining is compute-intensive, yet many tokens contribute marginally to learning, resulting in inefficiency. We introduce Efficient Selective Language Modeling (ESLM), an online, risk-aware batch selection algorithm that improves training efficiency and distributional robustness. ESLM operates in two phases: (i) instance-level selection via a shallow early-exit model pass that computes proxy per-instance statistics (e.g., loss or entropy) and retains data points using value-at-risk thresholding; and (ii) loss shaping with token-level selection via risk-aware thresholding on per-token scores. This data-centric mechanism reshapes the training objective, prioritizing high-risk tokens and eliminating redundant gradient computation. Framed as a bilevel game, the model competes with a masking adversary selecting worst-case token subsets, recovering conditional value-at-risk minimization, and linking selective pretraining to distributionally robust optimization. We extend our approach to Ada-ESLM, which adaptively tunes selection confidence during training. Experiments on GPT-2 pretraining show that ESLM substantially reduces FLOPs while matching or improving perplexity and downstream performance, scales across model sizes and dataset mixtures, and integrates naturally with knowledge distillation.
URL: https://openreview.net/forum?id=vcynsg6rMN
---
Title: A Behavior-first approach to Multi-Intention Inverse Reinforcement Learning
Abstract: Inverse Reinforcement Learning (IRL) seeks to infer reward functions from expert demonstrations. When demonstrations come from multiple experts with different intentions, the problem becomes Multi-Intention IRL (MI-IRL).
MI-IRL methods generally follow a reward-first paradigm: they ask which reward, if followed, could have generated each trajectory. This leads to trajectory similarity being based on reward likelihood rather than on the relationships between behaviors. In deep generative MI-IRL methods, this paradigm further couples behavior clustering and reward learning, making such methods dependent on prior knowledge of the true number of modes $K^*$. This limits adaptability to unseen behaviors, and restricts the analysis to learned rewards rather than relationships across behaviors. In contrast, we approach MI-IRL from a different standpoint. We propose Contrastive Multi-Intention IRL (CoMI-IRL), a transformer-based unsupervised behavior-first framework: it first learns a representation of trajectory dynamics that captures intrinsic behavioral similarity, then clusters trajectories in that space, and finally learns a reward for each cluster.
Our experiments show that CoMI-IRL outperforms existing approaches without needing $K^*$ or labels, while allowing for visual interpretation of behavior relationships and adaptation to unseen behavior without full retraining.
URL: https://openreview.net/forum?id=42kUyc9KGs
---
Title: Watch Before You Answer: Learning from Visually Grounded Post-Training
Abstract: It is critical for vision-language models (VLMs) to comprehensively understand visual, temporal, and textual cues. However, despite rapid progress in multimodal modeling, video understanding performance still lags behind text-based reasoning. In this work, we find that progress is even worse than previously assumed: commonly reported long video understanding benchmarks contain 40-60% of questions that can be answered using text cues alone. Furthermore, we find that these issues are also pervasive in widely used post-training datasets, potentially undercutting the ability of post-training to improve VLM video understanding performance. Guided by this observation, we introduce VidGround as a simple yet effective solution: using only the actual visually grounded questions without any linguistic biases for post-training. When used in tandem with RL-based post-training algorithms, this simple technique improves performance by up to 6.2 points relative to using the full dataset, while using only 69.1% of the original post-training data. Moreover, we show that data curation with a simple post-training algorithm outperforms several more complex post-training techniques, highlighting that data quality is a major bottleneck for improving video understanding in VLMs. These results underscore the importance of curating post-training data and evaluation benchmarks that truly require visual grounding to advance the development of more capable VLMs.
URL: https://openreview.net/forum?id=aopNknzMRY
---
Title: EveryGraph: Task-Agnostic Graph Neural Architectures
Abstract: Graph Neural Networks (GNNs) have demonstrated remarkable performance across a range of applications. However, their generalization across diverse tasks and datasets remains a fundamental challenge. Inspired by the success of foundation models in language, vision, and audio domains, recent efforts have sought to bring similar capabilities to graph-based learning.
We introduce EveryGraph, a task-agnostic GNN architecture that generalizes across more than 20 benchmark datasets spanning link prediction, node classification, and graph classification. EveryGraph introduces a novel method for constructing a unified representation space. Unlike traditional embedding techniques that may lose local context, our approach explicitly preserves the topological structure of the data, ensuring that intrinsic inter-node dependencies are accurately mapped into the latent space. Notably, EveryGraph is trained in a self-supervised regime on a single dataset and achieves this broad generalization by utilizing an in-context like downstream task adaptation.
URL: https://openreview.net/forum?id=htazhOL92a
---
Title: Regular Fourier Features for Nonstationary Gaussian Processes
Abstract: Simulating a Gaussian process requires sampling from a high-dimensional Gaussian distribution, which scales cubically with the number of sample locations. Spectral methods address this challenge by exploiting the Fourier representation and treating the spectral density as a probability distribution suitable for Monte Carlo approximation. Although this probabilistic interpretation is valid for stationary processes, it is overly restrictive for the nonstationary case, where spectral densities are generally not probability measures. We propose regular Fourier features for harmonizable processes to avoid this limitation. Our method discretizes the spectral representation directly, preserving the correlation structure among spectral weights without requiring probability assumptions. Under a finite-spectral-support assumption, this yields an efficient low-rank approximation that is consistent and positive semi-definite by construction. When the spectral density is unknown, the framework extends naturally to kernel learning from data. We demonstrate the method on locally stationary and harmonizable mixture kernels, the latter with a complex-valued spectral density, and apply the kernel-learning extension to real and synthetic data.
URL: https://openreview.net/forum?id=2eZhxVDAhR
---
Title: NOHA: A Diacritic-Aware Rule-Based Morphophonological Tokenizer for Arabic
Abstract: Arabic is a morphologically rich language whose complex phonological and morphological structure presents fundamental challenges for statistical tokenization methods used in large language model (LLM) pre-training. Although statistical methods such as Byte-Pair Encoding (BPE) are widely used, they are not optimally suited to morphologically rich languages such as Arabic, where frequency-based subword discovery ignores phonological and morphological boundaries and produces inconsistent segmentation of diacritized text. This paper presents NOHA, a fully deterministic diacritic-aware Arabic tokenizer that replaces statistical subword discovery with a rule-based chunking algorithm based on Arabic morphological and phonological structure. NOHA applies clitic preservation and diacritic-driven syllabic chunking to produce subword units that reflect the natural sound and grammatical structure of Arabic, without any statistical training. Despite producing approximately $1.49\times$ more tokens per sequence than BPE (and thus observing fewer unique training documents under identical compute and hyperparameter conditions), a GPT-2-scale model trained with NOHA achieves higher average accuracy across 38 matched-compute evaluation checkpoints on all three targeted Arabic linguistic benchmarks: $+2.42\%$ on GAT Reading Comprehension, $+2.65\%$ on the three Arabic language subsets of ArabicMMLU, and $+2.54\%$ on the Teacher-Authored Arabic Grammar benchmark (TAAG, $n{=}171$), a 171-question evaluation set developed and validated by a certified Arabic language teacher for this study. On general knowledge benchmarks, BPE leads by $0.26\%$ on a weighted average across 18 ArabicMMLU subsets, indicating that NOHA preserves general knowledge capacity while improving linguistic understanding. BPE achieves lower bits-per-byte at all 38 matched-compute evaluation checkpoints, with a mean compression advantage of $+2.87\%$. Both models perform near the $25\%$ random baseline on the Nahw-MCQ Arabic grammar benchmark, the remaining GATLc subtasks (Verbal Analogy, Sentence Completion, Contextual Error, and Semantic Association), and most mathematical reasoning subtasks, confirming that model scale rather than tokenizer design is the limiting factor for these tasks. For mathematical reasoning subtasks, limited exposure to mathematical content in the training corpus is an additional contributing factor. These findings show that linguistically grounded tokenization yields measurable improvements on Arabic linguistic tasks at an identical computational budget, and that compression efficiency and linguistic understanding are dissociable properties of Arabic tokenizer design.
URL: https://openreview.net/forum?id=Vyc4Ormkqp
---
Title: Revisiting continual learning through the unsupervized learning of dynamic contexts
Abstract: Continual learning from non-stationary data streams remains a fundamental challenge for autonomous agents. We revisit continual learning by proposing a formal framework that addresses the dual challenge of context alternation and within-context drift. We illustrate this on context-aware anomaly detection, where detecting anomalies alone is insufficient: the system must determine whether detected anomalies are hazardous within specific, evolving contexts. To address this, we introduce a benchmark with alternating contexts and smooth distribution drifts, and propose a solution leveraging topology-preserving vector quantization (GNG-T) to simultaneously track context-specific representations and register so-called contexts. Contexts encode both a structural characterization of the current distribution and
context-dependent classifiers. Experimental results demonstrate that explicit context recognition significantly improves recovery after context switches and enables effective anomaly hazard classification, outperforming approaches that continuously train a single classifier.
URL: https://openreview.net/forum?id=vT6Arrk5PK
---
Title: Controlling LLM Reasoning Behavior via Guided Decoding
Abstract: Large Language Models (LLMs) are increasingly being used to simulate human behavior in applications such as educational technology, user modeling, and human–AI interaction. However, LLMs often struggle to realistically simulate diverse reasoning behaviors. Instead, they tend to generate similar reasoning patterns even when explicitly prompted to emulate users with different reasoning behaviors.
In this work, we identify this phenomenon as the \textit{Reasoning Prior Bias}, where generation is dominated by a strong prior over certain reasoning patterns. To address this issue, we propose \textbf{Guided Decoding}, a decoding-time method for dynamically controlling the reasoning behavior of LLMs during inference. Our approach fuses token-level logits from two sources: a reference prompt that specifies the model’s primary reasoning behavior, and one or more guidance prompts that introduce alternative reasoning tendencies. By adjusting the contribution of these signals during decoding, Guided Decoding enables fine-grained control over the mixture of reasoning behaviors in generated responses. Experiments on multiple question answering benchmarks demonstrate that our method not only modulates task accuracy but also systematically alters the reasoning process.
URL: https://openreview.net/forum?id=EGbYJuUrFM
---
Title: The Scaling Laws of Classification Networks: Insights from Adaptive Exact Average Density Approximation
Abstract: Our main goal is to establish a theoretical foundation for the scaling laws that aligns with the empirical scaling laws observed in deep neural networks (DNNs). Under the assumption that the boundary of the target classification function is a semi-algebraic set, we show that the generalization error bound follows scaling laws for large networks. The rate of scaling with respect to sample size is intrinsically linked to the effective dimension of the data manifold, independent of the specific network model or learning algorithm. The scaling law with respect to the number of parameters, however, varies across learning methods and architectures — a variability captured by the \textit{box-model dimension}, which measures how the number of model parameters grows as the radius of covering balls decreases. As a secondary contribution, we show that modular network architectures exhibit a discrete scale-invariance property that enables the prediction of generalization error for larger models from smaller ones, and we empirically demonstrate this on ResNet, VGG, and Inception models trained on ImageNet, CIFAR-100, and Udacity.
URL: https://openreview.net/forum?id=Z4EK9gRnLr
---
Title: Greedy Decoding Is Not Precision-Invariant: Cross-Precision Output Divergence in LLM Inference
Abstract: Greedy decoding from large language models is commonly treated as deterministic. We show it is not precision-invariant: the same model, prompt, and decoding algorithm produce different outputs in BF16 versus FP16 on identical hardware. This affects 59-82% of prompts across six models (1.1B-12B parameters, four families) and three benchmarks; a single token flip cascades into a fully divergent generation. We develop an error-propagation theory whose central prediction is counterintuitive: 22 layers of accumulated body error are irrelevant to whether a token flips. The outcome depends entirely on a single scalar, the top-two logit margin at the lm_head. The theory makes five falsifiable predictions, including that applying more FP32 compute (broader scope) makes agreement worse, and all five are confirmed. The one effective intervention, selective FP32 lm_head recomputation, triggered only when the margin falls below a threshold, delivers +22-36pp exact agreement at <4% latency overhead. We map the applicability boundary across six models and four batch sizes, identifying training-time precision stability as the determining factor.
URL: https://openreview.net/forum?id=QDOKyg7a5e
---
Title: GenReview: A Large-scale Dataset of AI-Generated (and Human-written) Peer Reviews
Abstract: How does the progressive embracement of Large Language Models (LLMs) affect scientific peer reviewing? This multifaceted question is fundamental to the effectiveness—as well as to the integrity—of the scientific process. Recent evidence suggests that LLMs may have already been tacitly used in peer reviewing, e.g., at the 2024 International Conference of Learning Representations (ICLR). Furthermore, some efforts have been undertaken in an attempt to explicitly integrate LLMs in peer reviewing by various editorial boards (including that of ICLR’25). To fully understand the utility and the implications of LLMs’ deployment for scientific reviewing, a comprehensive relevant dataset is strongly desirable. Despite some previous research on this topic, such dataset has been lacking so far.
We fill in this gap by presenting Gen-Review, the hitherto largest dataset containing LLM-written reviews. Our dataset includes 81K reviews generated for all submissions to the 2018–2025 editions of the ICLR by providing the LLM with three independent prompts: a negative, a positive, and a neutral one. Gen-Review is also linked to the respective papers and their original reviews, thereby enabling a broad range of investigations. To illustrate the value of Gen-Review, we explore several intriguing research questions, namely: if LLMs favor accept-class ratings (they tend to do so) if LLM-written reviews can be automatically detected (so far, they can); if LLMs can rigorously follow reviewing instructions (not always) and whether LLM-provided ratings align with decisions on paper acceptance or rejection (holds true only for accepted papers). Gen-Review can be accessed at the following link: https://anonymous.4open.science/r/gen_review/
URL: https://openreview.net/forum?id=OMPsEQGLoj
---
Title: Multimodal Concept Bottleneck Models
Abstract: Concept Bottleneck Models (CBMs) enhance the interpretability of deep neural networks by aligning the features extracted from images with human-interpretable concepts. However, existing CBMs rely on a linear classifier, which not only restricts them to a fixed set of predefined classes, but also allows the model to bypass concept activations and compensate with non-concept predictive cues. In this paper, we propose Multimodal Concept Bottleneck Model (MM-CBM) to address these issues and extend CBMs into CLIP structure. MM-CBM utilizes dual Concept Bottleneck Layers (CBLs) to align both the image and text embeddings into interpretable features. This allows us to perform new vision tasks like zero-shot classification or image retrieval in an interpretable way. Compared to existing methods, MM-CBM achieves 23.28% accuracy improvement on average across four standard benchmarks. Our method maintains high accuracy, staying within 5% of black-box model performance under finetuned cases while offering greater interpretability.
URL: https://openreview.net/forum?id=wBagHznSw7
---
Title: Networked Communication for Decentralised Agents in Mean-Field Games
Abstract: Methods like multi-agent reinforcement learning struggle to scale with growing population size. Mean-field games (MFGs) are a game-theoretic approach that can circumvent this by finding a solution for an abstract infinite population, which can then be used as an approximate solution for the $N$-agent problem. However, classical MFG algorithms usually only work under restrictive conditions. We address this by introducing networked communication to MFGs, focusing on settings that use a single, non-episodic run of $N$ decentralised agents to simulate the infinite population, as is likely most reasonable in real-world deployments. Within this regime we situate our proposed networked algorithm against two pre-existing learning architectures - centralised and independent, which differ in how the $N$ agents share information during learning - and prove that the sample guarantees of our networked algorithm lie between those of these alternatives, with the exact position determined by the network structure and the number of communication rounds. These sample guarantees do not, however, translate into practical convergence times for any of the three theoretical algorithms. We thus contribute practical enhancements to all three algorithms, allowing us to present their first empirical demonstrations. We then show that in practical settings where the theoretical hyperparameters are not observed, giving fewer loops but poorer estimation of the Q-function, our communication scheme still respects the earlier theoretical ordering: under matched hyperparameters it considerably accelerates learning over the independent algorithm (which shows little improvement in reward within the training budget), and often performs comparably to the centralised algorithm while removing the latter's restrictive assumption. We provide ablations and additional studies showing that our networked approach is also more robust than both alternatives to update failures and to changes in population size.
URL: https://openreview.net/forum?id=U3EjlNgyk8
---
Title: Online Distributional Prediction via Latent Cluster Geometry Under Drift and Corruption
Abstract: Online learning in non-stationary streams is often formulated as tracking a point estimate, but many applications require predicting the full data-generating distribution. We study online distributional prediction under drift and adversarial corruption. Our approach represents each candidate law through a latent cluster geometry: a variable-size configuration of centers that organizes probability mass and induces a predictive distribution. A Gibbs quasi-posterior over these configurations yields an online predictor by posterior averaging, and the resulting variable-dimensional posterior can be sampled with reversible-jump MCMC. The method therefore avoids specifying a parametric streaming law while retaining a structured latent space for uncertainty, regularization, and comparison. We evaluate performance by cumulative Wasserstein-1 regret against the time-varying true law. The analysis separates two effects: corruption perturbs the loss-based posterior update, whereas drift makes long-horizon posterior memory stale. We address the latter with a restarted variant that temporally localizes the same quasi-Bayesian update. The resulting high-probability bounds decompose into a PAC-Bayesian complexity term, a corruption-sensitive posterior perturbation term, and a dynamic optimal-transport term driven by \(A_T^{\mathrm{OT}}=\sum_{t=2}^T W_2^2(p_{t-1}^*,p_t^*)\). Under bounded support, stable latent geometry, predictive-map regularity, oracle realizability, localized restart windows, sublinear transport action, and sublinear corruption budget, the restarted predictor achieves sublinear cumulative Wasserstein regret. These guarantees require no parametric model for the stream, drift mechanism, or corruption process.
URL: https://openreview.net/forum?id=DlL73vEtwX
---
Title: Learning Rate Transfer and Feature Learning Across Depth for Constrained Spectral Optimizers: Complete Scion
Abstract: Hyperparameter transfer across model depth is instrumental in training deep models, as it enables learning rates calibrated on shallow networks to be effectively applied to deeper architectures. The recent CompleteP framework establishes a $1/L$ residual scaling rule that demonstrates this for Adam-family optimizers. We study whether similar transfer extends to constrained optimization methods, specifically Frank-Wolfe and projected non-Euclidean normalized steepest descent, where updates are determined by linear minimization oracles. We derive scaling rules for these methods which confirm that the $1/L$ residual scaling applies, although the arguments differ from the Adam-family optimizers. Our analysis supports depth-wise transfer and width-wise transfer of the learning rate for the Scion algorithm while ensuring feature learning and avoiding the kernel-regime. In experiments on NLP transformers (nanoGPT on FineWeb), vision transformer (DeiT on ImageNet), and vision MLPs (MLP-Mixer on CIFAR-10), we observe learning rate transfer across depth and feature learning for appropriately chosen residual scaling.
URL: https://openreview.net/forum?id=TGiJpGVPNA
---
Title: Is the Prediction Set Size Well-Calibrated? A Closer Look at Uncertainty in Conformal Prediction
Abstract: Given its flexibility and low computation, conformal prediction (CP) has become one of the most popular uncertainty quantification methods in recent years. In deep classifiers, CP will generate a prediction set for a test sample that satisfies the $(1-\alpha)$ coverage guarantee. The prediction set size (PSS) is then considered a reflection of the predictive uncertainty. However, it is unknown whether the predictive uncertainty of CP is aligned with its predictive correctness, which is an imperative property for predictive uncertainty. This work answers this open question by investigating the uncertainty calibration of CP in deep classifiers. We first give a definition for the uncertainty calibration of CP by building a connection between PSS and prediction accuracy and then propose a calibration target for CP based on a theoretical analysis of the predictive distributions. Given this defined CP calibration, we present an empirical study on several classification datasets and reveal their weak calibration of CP. To strengthen the calibration of CP, we propose CP-aware calibration (CPAC), a bi-level optimization algorithm, and demonstrate the effectiveness of CPAC on several standard classification datasets by testing models including ResNet, Vision Transformer and GPT-2.
URL: https://openreview.net/forum?id=lv4gVtbve3
---
Title: Alert: Learning Trigger Functions for Early Classification of Time Series using Deep-RL
Abstract: Early Classification of Time Series (ECTS) is vital in fields like industrial monitoring and medical triage, where quick and accurate predictions are essential. The core challenge lies in the trigger function, which decides when to make a prediction, independently from the classifier itself. Most existing methods rely on handcrafted rules, but can data-driven approaches outperform them? This paper introduces Alert, a Deep RL framework that learns trigger functions from any state representation. Systematic comparisons on 30 datasets show that the design of the state space significantly influences performance. Building on this, we propose Alert+, a simple yet effective variant that consistently outperforms traditional methods in balancing accuracy and delay, while significantly reducing empirical risk. Alert and Alert+ are released to support reproducible research and practical applications.
URL: https://openreview.net/forum?id=J7EIFs4sjx
---
Title: Tuning without Peeking: Provable Generalization Bounds and Robust LLM Post-Training
Abstract: Gradient-based optimization is the workhorse of deep learning, offering efficient and scalable training via backpropagation. However, exposing gradients during training can leak sensitive information about the underlying data, raising privacy and security concerns such as susceptibility to data poisoning attacks. In contrast, black box optimization methods, which treat the model as an opaque function, relying solely on function evaluations to guide optimization, offer a promising alternative in scenarios where data access is restricted, adversarial risks are high, or overfitting is a concern.
This paper introduces BBoxER, an evolutionary black-box method for LLM post-training that induces an information bottleneck via implicit compression of the training data.
Leveraging the tractability of information flow, we provide non-vacuous generalization bounds and strong theoretical guarantees for robustness to data poisoning attacks and extraction attacks, while ensuring privacy by design.
In experiments with LLMs, we demonstrate empirically that black-box optimization methods—despite the scalability and computational challenges inherent to black-box approaches—are able to learn, showing how a few iterations of BBoxER improve performance, generalize well on a benchmark of reasoning datasets, and are robust to membership inference attacks. This positions BBoxER as an attractive add-on on top of gradient-based optimization, offering suitability for deployment in restricted environments while also providing non-vacuous generalization guarantees.
URL: https://openreview.net/forum?id=H5iA8BjtzI
---
Title: Beyond Size: How Gradients Shape Pruning Decisions in Large Language Models
Abstract: Large Language Models (LLMs) with billions of parameters are prime targets for network pruning, which removes some model weights without hurting performance. Prior approaches such as magnitude pruning, SparseGPT, and Wanda either concentrated solely on weights or integrated weights with activations for sparsity. However, they overlooked the informative gradients derived from pretrained LLMs. In this paper, we present a novel sparsity-centric pruning method for pretrained LLMs, termed ${\bf G}$radient-${\bf b}$ased ${\bf L}$anguage ${\bf M}$odel ${\bf P}$runer (GBLM-Pruner). GBLM-Pruner leverages the first-order term of the Taylor expansion, operating in a training-free manner by harnessing properly normalized gradients from a few calibration samples to determine the pruning metric, and substantially outperforms competitive counterparts like SparseGPT and Wanda in multiple benchmarks. Intriguingly, by incorporating gradients, unstructured pruning with our method tends to reveal some structural patterns, which mirrors the geometric interdependence inherent in the LLMs' parameter structure. Additionally, GBLM-Pruner functions without any subsequent retraining or weight updates to maintain its simplicity as other counterparts. Extensive evaluations on LLaMA-series, Qwen-Series and Deepseek models across zero-shot, STEM-Reasoning and Coding benchmarks show that GBLM-Pruner surpasses magnitude pruning and Wanda by significant margins and performs competitively against SparseGPT which involves intensive weight updates. We further extend our approach on Vision Transformer and Vision Language Models. Our code is available in the supplementary material.
URL: https://openreview.net/forum?id=TfuJfNemT2
---
Title: Relaxing Pointwise Imitation in Knowledge Distillation with Mean-Projection Matching
Abstract: Logit-based knowledge distillation (KD) typically trains a student by matching the teacher's predictive distribution. This pointwise target transfers useful soft information, but can also pass along miscalibrated, biased, or wrong teacher predictions. We revisit the supervision object and propose Mean Projection (MP), a distillation objective on the square-root probability sphere. For each gold-label subset in a mini-batch, MP decomposes the empirical mean vector of teacher predictions into a direction and a length, where the length is the average of sample-level scalar projections onto that direction. MP uses two bounded matching terms: one aligns the subset-level mean direction, and the other aligns these sample-level projections. This preserves the subset summary without forcing the student to match the full teacher distribution pointwise; in the single-sample limit, each term reduces to the squared Hellinger distance. Empirically, a natural language inference (NLI) diagnostic motivates the imperfect-teacher setting. On fine-grained classification, MP gives clear gains on visually similar datasets, where it matches the strongest logit-distillation baseline on accuracy while showing an expected calibration error (ECE) advantage among distillation methods. Mechanism analysis shows that MP predicts the gold label on teacher-wrong samples 2.9 percentage points more often than KD and copies teacher mistakes 4.1 percentage points less often. Further analysis traces these effects primarily to the square-root geometry, with subset relaxation contributing additional inheritance gains.
URL: https://openreview.net/forum?id=QhaAwWjtXa
---
Title: A new look at the Riemannian manifold \\ of positive definite matrices
Abstract: The Riemannian manifold of positive definite matrices endowed with the affine-invariant metric is an important tool in machine learning. The distance between two points P and Q on such a manifold is known to be a function of their generalized eigenvalues. We show that geodesics from P to Q, including the geometric mean, the logarithmic and exponential maps of Q at base-point P as well as the parallel transport of another point from P to Q, can be all expressed as a function of their generalized eigenvalues and eigenvectors. Thus, all important Riemannian manipulations on the manifold of positive definite matrices can be obtained by computing a joint diagonalizer. As compared to known expressions, those we derive for the geodesic, the logarithmic map, the exponential map and the parallel transport are simpler and lead to a unified conceptual and computational procedure, which is advantageous both in terms of speed and memory allocations. Backpropagation rules for applying this simplified framework in Riemannian neural networks are given.
URL: https://openreview.net/forum?id=txJmrkxLxk
---
Title: Beyond Neural Collapse: Task-Intrinsic Geometry Governs Neural Representations in Modular Arithmetic
Abstract: While neural collapse (NC) predicts that a $K$-class-balanced classifier should organize terminal representations as a $(K-1)$-dimensional simplex equiangular tight frame (ETF), modular addition consistently enters a different regime: networks compress to a two-dimensional cyclic geometry in which both classifier weights and token embeddings lie on circles. We refine the explanation of this phenomenon in three directions. First, we formalize a layerwise non-uniform training mechanism: downstream classifier weights are driven by dense cross-entropy gradients into a rank-2 equiangular configuration before upstream embeddings fully reorganize, and once this classifier plane forms, backpropagated feature gradients constrain embedding motion to the same plane while weight decay suppresses orthogonal components. Second, after this subspace locking, the induced in-plane dynamics admit an entropy-regularized transport interpretation on $S^1$; combined with modular-addition labels, this reduces embedding formation to phase alignment, whose minimizers are single-frequency characters of $\mathbb{Z}/P\mathbb{Z}$ and hence equal-angle points on a circle. Third, we quantify why this solution prevails over NC: a simplex ETF gains only an $O(1)$ advantage in cross-entropy, whereas the cyclic rank-2 solution enjoys a $\Theta(K)$ advantage under Schatten or weight-decay surrogates, yielding a critical threshold $\lambda_{\mathrm{crit}} = \Theta(1/K)$. Our results explain both why classifier weights move first and why embeddings subsequently align with them, showing that grokking on modular arithmetic is governed not by maximal separation alone but by a task-structured trade-off between separation, symmetry, and complexity.
URL: https://openreview.net/forum?id=ve7Iq6JQ6G
---
Title: Prior-Informed Flow Matching for Graph Reconstruction
Abstract: We introduce Prior-Informed Flow Matching (PIFM), a conditional flow model for graph reconstruction. Reconstructing graphs from partial observations remains a key challenge; classical embedding methods often lack global consistency, while modern generative models struggle to incorporate structural priors. PIFM bridges this gap by integrating embedding-based priors with continuous-time flow matching. Grounded in a permutation equivariant version of the distortion-perception theory, our method first uses a prior, such as GraphSAGE or node2vec, to form an informed initial state of the adjacency matrix based on local information. It then applies rectified flow matching to refine this estimate, transporting it toward the true distribution of clean graphs and learning a global coupling. Experiments on different datasets demonstrate that PIFM consistently enhances classical embeddings, outperforming them and state-of-the-art generative baselines in reconstruction accuracy.
URL: https://openreview.net/forum?id=9sEWrLXGPp
---
Title: A Controlled Synthetic Benchmark for Educational Aspect-Based Sentiment Analysis
Abstract: Large language models are increasingly used to synthesize labeled training data where annotation is scarce, but the generated labels are often unfaithful to the text, and it is difficult to tell whether a synthetic benchmark carries genuine learnable signal or merely reproduces label frequencies. We study both problems in educa-tional aspect-based sentiment analysis (ABSA), a setting where real aspect-labeled feedback is private and costly to annotate, with a methodology that applies beyond it. We release a controlled corpus of 10,000 synthetic course reviews over a 20-aspect pedagogical schema, generated by sampling supervision targets separately from nuance attributes, so the same labels recur across varied course contexts and styles. We then introduce a faithfulness-aware pipeline: a cost-matched LLM audit scores how well each declared label is supported by the text, and that score both fil-ters the training data and is itself validated across independent LLM judges and against human labels (Cohen's kappa 0.56 on two annotated corpora). Three con-trols establish that the corpus carries learnable signal rather than label priors: per-muting the labels collapses detection to the trivial floor (0.182 versus 0.276 micro-F1), accuracy scales monotonically with training size (0.183 to 0.285), and restrict-ing to faithfully labeled rows raises the ceiling. Evaluated by transfer to real human-annotated feedback (the unbiased metric), faithfulness-aware row filtering lowers transferred sentiment error on Herath across two architectures (paired 95% boot-strap CI excludes zero), and a lowest-faithfulness control collapses transfer on both real benchmarks (Herath and EduRABSA). Synthetic-only training recovers about 60% of a real-trained model, and synthetic pre-training followed by real fine-tuning exceeds real-only training. The audit-filter-validate recipe is a general tool for quali-ty-controlling LLM-supervised data; the corpus and code are released.
URL: https://openreview.net/forum?id=tmNAKpmSaF
---
Title: InquiTree: Evaluating AI Agents in the Scientific Inquiry Loop with Paper-Derived Research Trees
Abstract: While LLM-based agents are increasingly used in scientific workflows, it remains unclear whether they are truly qualified for the dynamic and uncertain process of discovery. Existing static evaluations often conflate genuine reasoning with rote memorization. We introduce InquiTree, a diagnostic environment that formalizes scientific inquiry as interactive Research Trees: directed acyclic graphs capturing the logical dependencies among hypothesis formulation, study design, result interpretation, and belief updating. Evaluating agents on a 30-paper test pool and releasing the open-access IT-18 subset, we identify two key limitations. First, agents exhibit an "Erosion of Marginal Capabilities": during long-horizon interactions, they develop "cognitive tunneling," where critical judgment and anomaly detection degrade relative to their intrinsic baselines. Second, performance drops on papers published after model training cutoffs, revealing a boundary between interpolation and extrapolation and suggesting that apparent competence is partly driven by parametric memory. These findings indicate that scaling context alone is insufficient for reliable AI scientists; stronger architectures or human oversight may be required to preserve critical evaluation and generalization.
URL: https://openreview.net/forum?id=esE2r2JRbD
---
Title: MLGLP: Multi-Scale Line-Graph Link Prediction Based on Graph Neural Networks
Abstract: Scale invariance is a critical property for deep models, enabling them to consistently recognize patterns across varying resolutions or granularities. In this paper, we leverage this principle by proposing a multi-scale link prediction approach based on Graph Neural Networks (GNNs). The proposed method, Multi-Scale Line-Graph Link Prediction (MLGLP), learns graph structure and extracts informative edge representations to address information loss, capture multi-scale structural information, and improve robustness when informative node features are limited or unavailable. This approach utilizes embedding vectors generated by GNNs from enclosing subgraphs. While expanding GNN layers can capture more intricate relations, it often leads to over-smoothing, where node representations become overly similar after many layers. To mitigate this issue, we propose constructing coarse-grained graphs at three distinct scales to uncover complex relations. To apply multi-scale subgraphs in GNNs without using pooling layers that lead to information loss, we convert each subgraph into a line-graph and reformulate the task as a binary node classification problem. The hierarchical structure facilitates exploration across three levels of abstraction, fostering deeper comprehension of the relationships and dependencies inherent within the graph. We perform extensive experiments on several well-known benchmarks and compare the results with state-of-the-art link prediction methods. The experimental results demonstrate the superiority of our proposed model in terms of average precision and area under the curve.
URL: https://openreview.net/forum?id=4EJVNEtYLf
---
Title: The Preference Centroid: Consensus Density Governs Output Dispersion in Aligned LLMs
Abstract: A growing literature reports that the same generative model equalizes outcomes for some users and amplifies differences for others. Sampling an aligned model repeatedly and embedding the completions, we summarize each output cloud by its dispersion—mean pairwise cosine distance—and report four findings. First, unconditioned dispersion is set by the consensus density of the task, rising from near-zero on factual prompts to wide on open-ended ones (Spearman ρ = 0.85, n = 20); the ordering replicates on a second, unrelated model (ρ = 0.84). Second, the construct is not circular: three held-out judge models, scoring only the prompt and never any output, predict the measured dispersion at ρ = −0.91, and a preliminary human expert panel tracks these judges at ρ = 0.74. Third, conditioning does not "sharpen" the model but relocates its sampling region: dispersion at the destination tracks the destination's own judge-scored consensus—falling below the unconditioned default for high-consensus relocations, rising above it for low-consensus ones—not the instruction's surface form. Fourth, a matched base-versus-instruct comparison on open weights shows the pretrained base already carries the gradient and alignment amplifies it, compressing the high-consensus end roughly 8.7-fold against 1.8-fold at the low end. Absolute magnitudes drift between sampling batches, so every claim rests on rank correlations and effect directions, which are stable across five fresh batches, two embedders, and four model families. These findings ground the equalizer/amplifier boundary at the distribution level rather than in posited task complexity; its further reading in terms of human expertise amplification is an interpretation the measurements motivate but do not themselves test.
URL: https://openreview.net/forum?id=6ukieTMBcG
---
Title: MixFlow 2: One-Stage MixFlow Training for Flow Matching
Abstract: MixFlow reduces the training--testing discrepancy by training the velocity prediction network at a training timestep $t$ using noisy teacher as training noisy data that correspond to training timestep $t$ and slowed timesteps (with higher noise). It is different from standard training: training noisy data is only from the teacher noisy data at the training timestep $t$. The original MixFlow approach adopts a two-stage training strategy: standard training first learns accurate instantaneous velocity prediction, and MixFlow post-training then uses noisy teacher forcing to handle the Slow Flow problem in multi-step discrete sampling. Direct application of original implementation to one-stage training, i.e., training from scratch, is not optimal. The reason is that training noisy data sampling is more from noisy teachers, and more capable of handling the Slow Flow problem, and less capable of learning instantaneous velocity prediction which is addressed by standard training with teacher forcing. We present a one-stage training approach, MixFlow 2, with the introduction of a novel sampling mechanism: the slowed timestep for forming the training noisy data is sampled from a symmetric triangular distribution, where intermediate slow timesteps are sampled more frequently. In addition, we observe that the prediction at high-noise training timesteps is less reliable, as high-noise timestep $t$ is less-frequently sampled. This influences guided sampling: guidance at high-noise timesteps is harmful. So we adopt guidance interval to remove the guidance at high-noise timesteps, maintaining the generation quality achieved without guidance. Empirical results show that MixFlow 2 improves the class-conditional and text-to-image generation performance and outperforms original MixFlow.
URL: https://openreview.net/forum?id=EK4L9zQHkH
---
Title: Speech-DRAME: Human-Aligned Evaluation for Speech Role-Play
Abstract: Role-play has become a key testbed for generative models, expanding from text-only dialogue to multimodal interaction. Extending role-play to speech captures prosody, emotion, and delivery, but also poses new evaluation challenges. Although audio large language models (ALLMs) used as zero-shot judges already exhibit non-trivial capability, current pipelines built on them suffer from three persistent issues: limited sensitivity to paralinguistic cues, collapse of diverse performance aspects into coarse scores, and reliance on synthetic speech references that fail to reflect real-world roles. We present Speech-DRAME, a unified framework that addresses these limitations through: (i) Speech-DRAME-EvalBench, an evaluation benchmark with bilingual human-annotated data, contrastive negative samples, and protocols for training and testing speech evaluation models (SEval), and (ii) DRAME-Eval, a fine-tuned SEval that builds on top of strong ALLM backbones to further improve agreement with human raters. Speech-DRAME distinguishes between two complementary evaluation strategies: Archetype Evaluation, a top-down approach measuring adherence to broad role archetypes, and Realism Evaluation, a bottom-up approach grounded in real human speech that emphasizes nuanced role quality. On the in-distribution test sets, where the spoken responses largely come from the same synthesis bank that supplies DRAME-Eval's training data, fine-tuning raises Pearson correlation with human ratings from 0.480 to 0.629 on archetypes and from 0.390 to 0.625 on realism over the strongest zero-shot ALLM-judge baseline (few-shot prompting reaches 0.495 and 0.432, respectively). On a separate real-recording test set drawn entirely from professional and amateur human acting, however, correlation remains modest: DRAME-Eval reaches an average Pearson correlation of 0.247, with the weakest results on identity- and trait-related dimensions. We report this synthetic-to-real gap as a measured open problem rather than a solved one, and treat closing it as the central direction for future work.
URL: https://openreview.net/forum?id=m9IP9ECrk2
---
Title: When Does Window Position Matter in Protein Language Models? A Controlled Benchmark and Cross-Architecture Taxonomy
Abstract: Protein language models (pLMs) are routinely applied to long proteins via fixed-length inference windows, so the same residue can be scored at the left edge, the center, or the right edge of its window. Using content-controlled probes, we ask whether and how this window position changes a pLM's behavior. We find a real but modest window-edge effect: bidirectional encoders (ESM-2, ESM C) are most faithful when a residue is centered and degrade toward the edges - an inverted-U, opposite the NLP "lost-in-the-middle" U and roughly an order of magnitude smaller. The effect is architecture-specific: causal models (ProGen2) instead depend monotonically on left context. We map it across six backbones along three axes (architecture, position encoding, structure-awareness). Mechanistically it is one-sided context-loss, not a terminal-token artifact; we find no decoder-style lost-in-the-middle and no truncation cliff - a controlled probe sliding a fixed-context residue across the training crop stays flat, consistent with rotary encodings extrapolating. Crucially, the effect is largely metric-level: it does not degrade three standard downstream tasks (variant-effect prediction, secondary structure, long-range contacts) when the window retains task-relevant context, the regime those tasks occupy by construction - but it does bite in the complementary regime, where cutting a long-range contact's partner to a different window makes the contact unrecoverable. We release a controlled benchmark, bias metrics, the taxonomy, the downstream scoping, and a training-free symmetric multi-crop remedy that recovers 68-80% of the edge loss without retraining.
URL: https://openreview.net/forum?id=kEpikXuZfK
---
Title: Attention to Mamba: A Recipe for Cross-Architecture Distillation
Abstract: State Space Models (SSMs) such as Mamba have become a popular alternative to Transformer models, due to their reduced memory consumption and higher throughput at generation compared to their Attention-based counterparts. On the other hand, the community has built up a considerable body of knowledge on how to train Transformers, and many pretrained Transformer models are readily available.
To facilitate the adoption of SSMs while leveraging existing pretrained Transformers, we aim to identify an effective recipe to distill an Attention-based model into a Mamba-like architecture. In prior work on cross-architecture distillation, however, it has been shown that a na\"ive distillation procedure from Transformers to Mamba fails to preserve the original teacher performance, a limitation often overcome with hybrid solutions combining Attention and SSM blocks.
The key argument from our work is that, by equipping Mamba with a principled initialization, we can recover an overall better recipe for cross-architectural distillation. To this end, we propose a principled two-stage approach: first, we distill knowledge from a traditional Transformer into a linearized version of Attention, using an adaptation of the \emph{kernel trick}. Then, we distill the linearized version into an adapted Mamba model that does not use any Attention block.
Overall, the distilled Mamba model is able to preserve the original Pythia-1B Transformer performance in downstream tasks, maintaining a perplexity of 14.11 close to the teacher's 13.86. To show the efficacy of our recipe, we conduct thorough ablations at 1B scale with 10B tokens varying sequence mixer architecture, scaling analysis on model sizes and total distillation tokens, and a sensitivity analysis on tokens allocation between stages.
URL: https://openreview.net/forum?id=uE0VwRVEyV
---
Title: Decentralized SGD with Controlled Disagreement Finds Flatter Minima
Abstract: Decentralized training is often regarded as inferior to centralized training because the consensus errors between workers are thought to undermine convergence and generalization. This work challenges this view by introducing decentralized SGD with Adaptive Consensus (DSGD-AC), which uses a time-dependent scaling mechanism to maintain consensus errors throughout the training. We show that adaptive consensus changes the stationary variance of disagreement modes by balancing two effects: it preserves consensus-error magnitude through weaker graph damping while still allowing curvature-dependent damping to shape the disagreement directions. This balance can produce a stronger Hessian-weighted loss-envelope penalty around the deployed model, even when normalized Hessian alignment is weaker than in standard DSGD. Empirical results on image classification show that DSGD-AC reaches flatter solutions and higher test accuracy than standard DSGD and even centralized SGD. Together, these results support consensus errors as a useful implicit regularizer and open a new perspective on the design of decentralized learning algorithms.
URL: https://openreview.net/forum?id=DisRuxcT57
---
Title: Decentralized Autoregressive Generation
Abstract: The decentralization of autoregressive generation has attracted considerable attention in recent years as a solution to scaling bottlenecks. However, despite promising empirical results, this paradigm currently lacks rigorous theoretical justification. In this work, we formally establish the theoretical equivalence between decentralized and centralized training. To achieve this, we adapt the Discrete Flow Matching framework for autoregressive generation, leveraging its inherent properties to demonstrate that global models naturally decompose into independent experts. Finally, we conduct extensive experiments across diverse multimodal benchmarks, empirically validating that decentralized training maintains competitive parity with standard centralized architectures.
URL: https://openreview.net/forum?id=omDa4FZ0NQ
---
Title: Energy-Based World Models for Controllable Text Generation with Frozen LLMs
Abstract: Large language models (LLMs) generate fluent text but, as monolithic systems, lack native mechanisms for fine-grained clamps on individual concepts (e.g., clamping a single attribute such as brand preference while holding everything else fixed). We propose an architectural separation grounded in form/meaning distinction: a frozen LLM provides linguistic competence (form), while a learned energy-based world model captures the latent belief structure (meaning) on which clamps act. Concretely, a Deep Boltzmann Machine (DBM) models the domain's latent co-occurrence structure, a lightweight adapter projects its mean-field beliefs into soft prompts, and a frozen Qwen2.5-0.5B-Instruct renders them as text. This split yields a falsifiable prediction: a form-optimized direct-MLP (No DBM) baseline should match or marginally beat the proposed model on form-level metrics (CE, SBERT), while meaning-level capabilities, including energy-based coherence scoring, predictively-valid clamping, and independent two-axis attribute control, should be exclusive to the DBM-based pipeline. We evaluate this on two Amazon review domains (Smartphones, Beauty) under a rigorous 3-seed protocol. On form metrics, the proposed model statistically dominates LoRA, full fine-tuning, CTRL, PPLM, and one-shot ICL in both domains ($p<10^{-5}$) and ties direct-MLP as predicted. On meaning-level metrics, it outperforms direct-MLP on every cross-domain $\times$ cross-metric cell: energy-based scoring recovers each domain's market structure while returning near-null scores on control pairs; coupling-aware clamping cuts Wasserstein distance to natural low-rating distributions by $23\%$ on Beauty; and two-axis clamps yield $1.6$--$3.4$ times larger content shifts than direct-MLP with sentiment held fixed. A generously tuned activation-steering baseline (CAA) matches the raw sentiment shift but is rejected on distributional match to natural reviews under RoBERTa, composes destructively under multi-attribute clamps, and entangles sentiment when steered on the content axis. Architectural separation thus offers a practical, theoretically grounded path toward controllable, interpretable, and model-consistent counterfactual text generation in structured domains.
URL: https://openreview.net/forum?id=pOIFHY4dOJ
---
Title: Frozen Features Often Match Fine-tuning: A Controlled Evaluation of Geospatial Foundation Models for Agricultural Field-Extent Segmentation
Abstract: Most published evaluations of geospatial foundation models (GeoFMs) fine-tune them end-to-end. We test whether this is necessary for agricultural field-extent segmentation from single-date Sentinel-2. In a single-recipe comparison across six countries, two GeoFMs, and ten matched training seeds, a frozen backbone with a trained decoder is TOST-equivalent to full fine-tuning at ε=0.02 AUROC in 9 of 12 region×model cells (Prithvi 5 of 6, TerraMind 4 of 6). Kenya is the consistent sparse-positive exception: full fine-tuning wins for both GeoFMs with CIs excluding zero. TerraMind India is also non-equivalent, but in the frozen-favored direction and with the CI extending past the equivalence margin. Cross-region transfer is model- and regime-dependent: on the full 30 directed transfer matrix a from-scratch U-Net is strongest (0.826 mean AUROC), both frozen GeoFMs trail it, and Prithvi frozen trails Prithvi full fine-tuning by 0.013 AUROC. On the non-Kenya 20-transfer subset, both frozen GeoFMs beat their full-FT counterparts, and TerraMind frozen is statistically indistinguishable from the U-Net. The frozen decoder trains about 110× fewer parameters than full fine-tuning.
We then examine two evaluation choices that affect the apparent GeoFM advantage. Replacing polygon-derived field-extent labels with ESA WorldCover “cropland” turns the task into a more spectrally separable land-cover problem, and comparing only against a per-pixel spectral baseline understates a task-specific supervised model. On identical pixels, a from-scratch U-Net reaches 0.89–0.98 AUROC, has higher point AUROC than a Prithvi frozen probe in five of six regions, and wins IoU in five of six in the ten-seed decoder comparison; India is the frozen-probe AUROC exception. PANGAEA reports a similar supervised-baseline pattern on AI4SmallFarms. Our claims are limited to single-date, per-pixel field-extent segmentation rather than parcel-boundary or instance delineation. We report the sparse-positive Kenya boundary case and provide a ten-point evaluation protocol.
URL: https://openreview.net/forum?id=qRXVTe1yYp
---
Title: Spectral Analysis of Molecular Features: When Richer Features Do Not Guarantee Better Generalization
Abstract: The spectral properties of feature embeddings offer critical insights into model generalization and representation quality. While deep learning models are widely used for molecular property prediction, kernel methods remain competitive in low-data regimes, yet their spectral behavior is largely unexplored.
We present the first comprehensive spectral analysis of kernel ridge regression across diverse representations—including molecular fingerprints (ECFP), pretrained transformers, graph neural networks, and 3D descriptors—evaluated on QM9 and 3 MoleculeNet benchmarks. Surprisingly, richer spectral features do not consistently yield better generalization performance, contradicting common representation heuristics used in self-supervised learning (SSL). Across 4 spectral metrics, only ECFP-based kernels show a strictly positive correlation with performance. Transformer and global 3D representations exhibit mixed behavior, whereas local 3D representations show consistently negative correlations. Truncation analysis further emphasizes this disparity: for local 3D representations on thermodynamic targets, fewer than 2\% of eigenvalues (and occasionally as few as 0.02\%) are needed to recover 95\% of performance, whereas ECFP and transformer kernels require significantly more. By demonstrating a strong dependence on both task and representation, our results challenge the heuristic that richer spectra inherently improve generalization, providing new guidance for evaluating representations in SSL and in label-limited scientific tasks.
URL: https://openreview.net/forum?id=teXV1OiQCo
---
Title: A Function-Centric Perspective on Flat and Sharp Minima
Abstract: Flat minima are strongly associated with improved generalisation in deep neural networks.
However, this connection has proven nuanced in recent studies, with both theoretical counterexamples and empirical exceptions emerging in the literature. In this paper, we revisit the role of sharpness in model performance and examine sharpness as a function-dependent property rather than as a direct indicator of poor generalisation. We conduct extensive empirical studies ranging from single-objective optimisation, synthetic non-linear binary classification tasks, to modern image classification tasks. In single-objective optimisation, we show that flatness and sharpness are relative to the function being learned: equally optimal solutions can exhibit markedly different local geometry. In synthetic non-linear binary classification tasks, we show that increasing decision-boundary tightness can increase sharpness even when models generalise perfectly, suggesting that sharpness is not reducible to memorisation alone. Finally, in large-scale experiments, we find that sharper minima often emerge when models are regularised (e.g., via weight decay, data augmentation, or SAM), and can coincide with better generalisation, calibration, robustness, and functional consistency. Our findings are consistent with a function-centric interpretation of minima geometry: flatness should not be treated as a universal proxy for generalisation, and sharper minima can reflect desirable inductive biases associated with improved predictive behaviour, calling for a function-centric reappraisal of minima geometry.
URL: https://openreview.net/forum?id=LV1ffpXXeV
---
Title: Granularity-Aware Class-Agnostic Object Counting
Abstract: Exemplar-based class-agnostic counters can count novel objects from a few examples, but they often ignore the granularity specified by those examples. Given exemplars of red apples, should the model count only red apples, or all visually related apples? Existing methods typically lack this control and tend to collapse fine-grained object groups into a parent category. We introduce granularity-aware class-agnostic counting and propose S-Seg, an exemplar-conditioned segmentation model controlled by a sensitivity parameter $\alpha$. Small $\alpha$ enforces strict exemplar matching, while large $\alpha$ expands the target to semantically related objects. The predicted mask then restricts existing counters to the intended target regions.
We train S-Seg from synthesized FSC-147 samples using only point supervision, without dense multi-granularity annotations. We also introduce FSC-FG, a benchmark with multiple fine-grained object groups per image. Experiments show that S-Seg consistently improves state-of-the-art counters on FSC-FG, reducing MAE by up to 54\%, with further gains when negative exemplars are used to resolve ambiguous target definitions.
URL: https://openreview.net/forum?id=pcpV7bYYqk
---
Title: $\pi$-CoT: Prolog-Initialized Chain-of-Thought\\ Prompting for Multi-Hop Question-Answering
Abstract: Chain-of-Thought (CoT) prompting significantly enhances large language models' (LLMs) problem-solving capabilities, but still struggles with complex multi-hop questions, often falling into circular reasoning patterns or deviating from the logical path entirely. This limitation is particularly acute in retrieval-augmented generation (RAG) settings, where obtaining the right context is critical. We introduce Prolog-Initialized Chain-of-Thought ($\pi$-CoT), a novel prompting strategy that combines logic programming's structural rigor with language models' flexibility. $\pi$-CoT reformulates multi-hop questions into Prolog queries decomposed as single-hop sub-queries. These are resolved sequentially, producing intermediate artifacts, with which we initialize the subsequent CoT reasoning procedure. Extensive experiments demonstrate that $\pi$-CoT significantly outperforms standard RAG and in-context CoT on multi-hop question-answering benchmarks.
URL: https://openreview.net/forum?id=gykgXffnE8
---
Title: Exact Convex Reformulations of Linear Neural Networks via Completely Positive Lifting
Abstract: We show that the training problem of a deep linear neural network under the squared loss admits an exact convex reformulation in a lifted space over a generalized completely positive cone. The reformulation has the same optimal value as the original nonconvex problem and is linear in the lifted variables, with all nonconvexity encoded in the cone constraint. Its ambient lifted dimension depends only on the input and output dimensions, independent of the network depth and the number of data points, and the bottleneck width enters only through scalar constraints. The construction proceeds by reducing the multilayer parameterization to a bilinear factorization, lifting it to a rank-constrained semidefinite program, expressing the rank constraint via a complementarity condition, and applying a completely positive lifting. While the resulting formulation is computationally intractable in general, it gives an exact conic representation of the nonconvexity induced by linear factorization and connects linear neural network training with copositive programming.
URL: https://openreview.net/forum?id=2l52hL2XrN
---
Title: COOT Accelerates Online Diffusion Policy Optimization from Confounded Offline Data
Abstract: Diffusion policies are powerful multimodal action generators, but online reinforcement learning with diffusion policies remains sample intensive when useful behaviors must be discovered through interaction alone. Offline-to-online learning can reduce this cost, yet practical logs are often biased by hidden execution regimes: unobserved factors may affect the logged action, the action actually executed by the environment, and the realized outcome. As a result, naive behavior cloning or offline value initialization can transfer spurious action-outcome correlations rather than reliable control knowledge. We propose COOT (Causally Motivated Offline-to-Online Transfer), a framework for accelerating online diffusion-policy optimization from biased but informative offline data. COOT uses observed proxy variables to build a proxy-informed reliability transfer bundle that estimates robust proxy-conditioned transfer value, bridge uncertainty, state potential, and behavior-support reliability. These signals selectively weight diffusion-policy pretraining, shape critic initialization, and regularize online candidate selection from both the current policy and a persistent offline prior. Under a same-data control with the identical offline dataset and pretraining budget, COOT improves over a uniform behavior-cloning warm start, showing that its gains are not explained by generic offline pretraining alone. Across five MuJoCo locomotion tasks and five random seeds, COOT also outperforms eight strong actor-critic, diffusion-policy, energy-based, Q-guided, and score-based online baselines in both peak return and sample efficiency. Ablations, proxy stress tests, and covariate-shift diagnostics further show that COOT's advantage comes from proxy-informed, uncertainty-aware, and support-aware transfer.
URL: https://openreview.net/forum?id=hbA8W0PsoK
---
Title: Exploring Extrinsic and Intrinsic Properties for Effective Reasoning with Code Interpreter
Abstract: Reasoning with a Code Interpreter (CI) has emerged as an effective paradigm for enhancing the reasoning capabilities of large language models (LLMs) through executable computation and iterative verification. Despite its growing adoption, the behavioral properties underlying effective code reasoning remain largely underexplored. In this work, we investigate code reasoning from two distinct perspectives inspired by prior studies of natural language reasoning: extrinsic properties, represented by crucial tokens, and intrinsic properties, represented by code-specific cognitive behaviors. Across multiple LLMs, we find that stronger CI reasoning models consistently exhibit a higher prevalence of crucial tokens and cognitive behaviors, particularly verification, backtracking, and backward chaining. Building on these observations, we examine how these properties can be leveraged during both inference and training. At inference time, appending code-specific crucial tokens improves performance on several reasoning capabilities, including mathematical, ordering, and optimization, while yielding limited benefits elsewhere. At training time, augmenting a state-of-the-art framework with code-specific cognitive behaviors improves supervised fine-tuning and reinforcement learning performance in two of three evaluated models. Further analysis shows that these behaviors reduce overthinking in incorrect responses and improve token efficiency, while also revealing factors that limit gains in a certain model. Our findings provide the first systematic characterization of effective reasoning with CI and demonstrate both the potential and limitations of leveraging key properties to improve CI-based reasoning.
URL: https://openreview.net/forum?id=CGGLbPh6qb
---
Title: Instruct, Don’t Rush to Fine-Tune! A Survey on the Art of Instructing Your AI for Desired Behaviors
Abstract: Despite the development of large-scale machine learning (ML) models with impressive capabilities, their widespread adoption remains limited by the difficulty in using them out of the box. How to effectively adapt these general-purpose models to application- or user-desired goals and behaviors has emerged as a central challenge. Achieving such reliable and domain-aligned performance often requires deliberate customization beyond simple prompt engineering or naive, expensive fine-tuning. We argue that ``smart instruction'' plays an important role in enabling efficient and effective adaptation. In this survey, we discuss emerging approaches for instructing large scale ML models such as multimodal large language models (LLMs), focusing on how choices in instructional structure affect robustness, generalization, and data efficiency. We introduce a taxonomy that characterizes these approaches as distinct modes of instruction and provide a synthesis of common principles, tradeoffs, and design patterns that emerge across domains, from large language models to agentic AI systems and physical AI. We organize this landscape into six complementary instructional forms: structural, in-context, knowledge-guided, spatial, temporal, and global, each capturing a distinct way of steering model behavior through reasoning structure, examples, evidence, geometry, procedural sequencing, or long-range context. We conclude with two case studies demonstrating how even practitioners without deep ML expertise can leverage this taxonomy to adapt modern ML systems in real-world settings.
URL: https://openreview.net/forum?id=MvCaoQGqaU
---
Title: 3D Primitives are a Spatial Language for VLMs
Abstract: Vision-language models (VLMs) exhibit a striking paradox: they can generate executable code that reconstructs a 3D scene from geometric primitives with correct object counts, classes, and approximate positions, yet the same models fail at simpler spatial tasks when directly answering natural-language questions on the same image. We show that 3D geometric primitives (cubes, spheres, cylinders, expressed in executable code) serve as a powerful intermediate representation for spatial understanding, and we exploit this insight through three contributions.
First, we introduce \textbf{\textsc{SpatialBabel}}, a benchmark that evaluates fourteen VLMs on primitive-based 3D scene reconstruction across six \emph{scene-code languages} (programming languages and declarative formats for expressing 3D primitive scenes), revealing that a model's object-detection F1 on identical images swings sharply with the target scene-code language: aggregated with class- and scale-fidelity into our Reconstruct Score, the best-versus-worst-language gap reaches $0.40$ (on a $0$--$1$ scale) within a single model.
Second, we propose \textbf{Code-CoT} (Code Chain-of-Thought), a training-free inference strategy that routes spatial reasoning through primitive-based code generation. Code-CoT improves the SpatialBabel-QA-Score by up to $+8.9$\% on primitive scenes and lifts real-photo CV-Bench-3D accuracy by up to $+5.9$\% for VLMs with strong coding capabilities.
Third, we propose \textbf{S$^{3}$-FT} (Self-Supervised Spatial Fine-Tuning), which self-supervisedly distills primitive spatial knowledge into general visual reasoning by parsing the model's own primitive-reconstruction code (generated in Three.js) into structured annotations and fine-tuning on the result, requiring \emph{no human labels and no teacher model}. Training on primitive images alone, S$^3$-FT improves Qwen3-VL-8B by $+5.5$ to $+10.3$\% on SpatialBabel-Primitive-QA, $+9.7$\% on CV-Bench-2D, and $+17$\% on HallusionBench; the recipe transfers across model families.
These results establish geometric primitives in executable code as both a diagnostic tool for VLM spatial understanding and a transferable spatial vocabulary.
We will release the benchmark, training data, checkpoints, and evaluation toolkit upon publication.
URL: https://openreview.net/forum?id=w7hUWDP4iD
---
Title: Squeeze3D: Extreme Neural Compression with Latent Space Bridging
Abstract: We propose Squeeze3D, a novel framework that leverages implicit prior knowledge learnt by existing pre-trained encoders and decoders to compress 3D data at extremely high compression ratios. Our approach bridges the latent spaces between a pre-trained encoder and a pretrained decoder model through trainable mapping networks. Any 3D asset represented as a mesh, point cloud, or a radiance field is first encoded by the pre-trained encoder and then transformed (i.e. compressed) into a highly compact latent code by a mapping network. This latent code can effectively be used as an extremely compressed representation of the mesh, point cloud, or radiance field. A mapping network transforms the compressed latent code into the latent space of a powerful generative model, the decoder of this generative model then recreates the original 3D asset (i.e. decompression). Squeeze3D is trained entirely on generated synthetic data and does not require any 3D datasets. The Squeeze3D architecture can be flexibly used with existing pre-trained 3D encoders and existing generative models. It can flexibly support different formats, including meshes, point clouds, and radiance fields. Our experiments demonstrate that Squeeze3D achieves compression ratios of up to 2187$\times$ for textured meshes, 58.5$\times$ for point clouds, and 619$\times$ for radiance fields while maintaining visual quality comparable to many existing methods. Squeeze3D only incurs a small compression and decompression latency since it does not involve training object-specific networks to compress an object. We provide extensive additional visual results at https://anonymous.4open.science/w/tmlr-s3d-website/ and code at https://anonymous.4open.science/r/tmlr-s3d-code.
URL: https://openreview.net/forum?id=XXYGHfqzvA
---
Title: Refining Causal Graphs Beyond Conditional Independence: Kernel Two-Sample Testing on Intervention Data
Abstract: The accuracy of causal effect estimation is governed by the accuracy of the underlying causal graph. Kernel-based conditional independence testing (KCIT), a standard approach to causal discovery, suffers from structural limitations that persist regardless of sample size. Edges obscured by latent confounding or collider bias go undetected, and under faithfulness violations, where direct and indirect effects cancel, edge detection becomes impossible in principle. We propose an integrated algorithm that overcomes these limitations by using an information source fundamentally different from conditional independence, the distributional shift in intervention data, measured via Maximum Mean Discrepancy (MMD). An MMD intervention effect map detects edges that KCIT misses and determines causal direction. Conditional MMD isolates direct effects by controlling for mediators, recovering edges under complete faithfulness violations across all seeds. Within the KCIT stage, we introduce quantum-inspired kernel joint encoding to improve sensitivity to multiplicative dependencies prevalent in physical systems. We formalize the conditions under which a physics simulator serves as a valid surrogate for Pearl's do-operator ($\mathcal{S}$-valid do-operator) and verify that three widely-used scientific simulators satisfy these conditions. On the same synthetic benchmark, the proposed method achieves an F1 of 0.673, outperforming IGSP (0.430) and UT-IGSP (0.400) by a wide margin. On three real-world scientific datasets (drug toxicity, ultrafast laser physics, and battery chemistry), our method correctly estimates positive causal effects for all outcome variables in settings where methods lacking directional information produce sign-reversed estimates due to overadjustment.
URL: https://openreview.net/forum?id=qy8DIAEGrU
---
Title: Concept Attractors in LLMs and their Applications
Abstract: Large language models (LLMs) often map semantically related prompts to similar internal representations at specific layers, even when their surface forms differ widely. We show that this behavior can be explained through Iterated Function Systems (IFS), where layers act as contractive mappings toward concept-specific Attractors. We leverage this insight and develop simple, training-free methods that operate directly on these Attractors to solve a wide range of practical tasks, including language translation, hallucination reduction, guardrailing, and synthetic data generation. Despite their simplicity, these Attractor-based interventions match or exceed specialized baselines, offering an efficient alternative to heavy fine-tuning, generalizable in scenarios where baselines underperform.
URL: https://openreview.net/forum?id=N9bm8nADFD
---
Title: MemoryVLN: Memory-Augmented Vision-Language Navigation
Abstract: Vision-Language Navigation (VLN) requires embodied agents to interpret natural language instructions and navigate through previously unseen complex environments. Despite recent progress with large-scale pretraining and multimodal foundation models, current VLN systems remain limited by shallow temporal modeling and insufficient spatial understanding, often processing only short observation sequences and lacking structured memory. Such deficiency severely hinders them from capturing critical information from earlier observations and building detailed representations with structured spatial contexts, unable to make reliable decisions. In this work, we present MemoryVLN, a memory-augmented VLN framework that explicitly models temporal, spatial, and trajectory information to enable long-horizon reasoning and comprehensive spatial understanding. It includes a temporal reasoning memory to adaptively distinguish short-term dynamics and long-term context through selective frame retention, a spatial intelligence memory to build detailed scene graphs with object entities and relationships, and a trajectory memory to visualize historical actions as top-down trajectory maps to encode holistic motion history. With information-intensive memory features, MemoryVLN demonstrates new state-of-the-art performance on R2R-CE and RxR-CE benchmarks by outperforming previous methods by 5%-10%, while simultaneously achieving high computing efficiency by consuming far fewer input tokens. It even largely outperforms previous models with panoramic observations by using RGB observations only. Moreover, experiments on cross-dataset generalization and enhanced spatial reasoning further validate the effectiveness of MemoryVLN in capturing long-term temporal dependencies and 3D scene understanding for embodied navigation.
URL: https://openreview.net/forum?id=vHb61DlKTh
---
Title: Training LLMs to Self-Refine via Iterative Preference Optimization
Abstract: Self-Refinement represents a model's ability to revise its own responses to produce improved outputs. However, our experiments show that current large language models (LLMs) exhibit limited Self-Refinement capability and often experience performance degradation during the refinement process. This suggests that simply prompting models to revise their previous outputs is insufficient. To address this limitation, we propose EVOLVE, a framework for reinforcing Self-Refinement through iterative preference training. We first investigate optimization methods during training to activate and strengthen the model's Self-Refinement ability. Specifically, our training objective encourages the model not only to generate high-quality responses from the original prompt, but also to transform suboptimal responses into preferred ones. At inference time, we further explore diverse generation strategies to effectively exploit this capability while supplying the necessary data for continued training. By jointly optimizing both training and inference, our framework enables the continual evolution of Self-Refinement, allowing models to more effectively refine their own responses. This creates a closed-loop process in which improved refinement ability leads to better generated data, which in turn further enhances the model through subsequent preference optimization. Experiments on instruction-following benchmarks show that the learned Self-Refinement ability from EVOLVE enables the Llama-3.1-8B base model to surpass both Llama-3.1-405B-Instruct and GPT-4o, achieving 62.3% length-controlled and 63.3% raw win rates on AlpacaEval 2, and 50.3% on Arena-Hard. It also generalizes effectively to out-of-domain reasoning tasks, improving performance on mathematical reasoning benchmarks such as GSM8K and MATH.
URL: https://openreview.net/forum?id=8uL4gF73Jm
---
Title: Evaluating transfer learning for treatment-effect estimation in small samples
Abstract: Estimating how treatment effects vary across individuals typically requires large samples, yet in many applied settings, such as small clinical cohorts, regional surveys, and field experiments, the available data are limited and treatment is not randomized. We ask whether transfer learning can relax this constraint by borrowing strength from a larger, related source dataset. Building on the Treatment-Agnostic Representation Network (TARNet; Shalit et al., 2017) and its transfer-learning extension (Aloui et al., 2023), we provide the first systematic evaluation of this framework in small-sample, non-randomized settings, through two simulation studies and an empirical application. The first simulation constructs the target by subsampling the source, which isolates target sample size and treatment-selection bias. Transfer reduces both error and bias whenever the target is small or violates ignorability, and the gains from source data. The second simulation introduces genuine source-target mismatch corresponding to the two mismatch terms in the transfer bound: covariate shift and changes in the causal mechanism. Transfer is robust to covariate shift and purely prognostic changes, but it degrades steadily once the treatment effect itself changes, the one difference that alters the quantity being estimated. Moreover, the Causal Inference Task Affinity (CITA) score is far more sensitive to covariate shift than to the effect changes that most threaten transfer. In an application to the India Human Development Survey, transfer stabilizes otherwise implausible small-sample estimates of the effect of household firewood collection on children's study time, aligning them in sign and magnitude with the larger source. We close with practical guidance for transfer learning in small-sample causal estimation.
URL: https://openreview.net/forum?id=CgWR47MKfc
---
Title: Reproducing INP-Former: Reconstructing Intrinsic Normal Prototypes for Industrial Anomaly Detection
Abstract: We reproduce the central claims of Luo et al. (2025), which introduces INP-Former, a
reconstruction-based method for multi-class industrial anomaly detection. The method ex-
tracts Intrinsic Normal Prototypes (INPs) from test images themselves, avoiding the need for
stored memory banks, and uses them to guide a decoder that reconstructs normal features.
We verify the main performance claims on MVTec AD and VisA using the authors’ released
code, extend the evaluation to MVTec AD 2, and conduct ablation studies on the loss
components, number of prototypes, input resolution, and backbone architecture. Beyond
reproduction, we propose several extensions: (1) LoRA-based parameter-efficient transfer
learning from Real-IAD to downstream benchmarks, (2) an analysis of how anomaly area
size affects detection performance, (3) an inference speed benchmark, (4) a lighting robust-
ness evaluation on MVTec AD 2, and (5) an epoch efficiency analysis showing convergence
behavior. Our reproduction supports the central claims: INP-Former achieves state-of-
the-art multi-class anomaly detection, and both the INP mechanism and the proposed loss
terms contribute meaningfully to performance. We additionally find that the optimal num-
ber of prototypes varies across datasets, that LoRA adaptation can recover most of the
fully-trained performance using only 5% of trainable parameters, that the model largely
converges by epoch 50, that small anomalies are the primary failure mode, and that the
method is robust to brightness changes but vulnerable to color shifts.
The code we used for reproducing the study can be found here: https://anonymous.4open.science/r/Repro-INPFormer-D6D1/README.md
URL: https://openreview.net/forum?id=wohg1BSS7m
---
Title: An Analysis of Safety Mechanisms for Constrained Reinforcement Learning with Gradient Boosted Regression Trees
Abstract: One aspect of trustworthy autonomous systems is the property of transparent decision-making,
one that neural network architectures struggle with. Gradient-Boosted Regression
Trees offer an alternative to neural networks as function approximators for Reinforcement
Learning (RL), but their integration with safety-constrained policy optimization methods
has not been studied systematically. We show that the structural properties of boosted
ensembles, in particular frozen learners and function-space rather than parameter-space
optimization, invalidate assumptions on which existing Safe RL algorithms rely. We identify
five architectural adaptations that are necessary for correct integration: explicit critic
orchestration, scale-consistent projection coefficients, independent ensemble allocation, an
undiscounted violation signal, and a bounded ensemble combined with joint freezing of continuous
policy parameters. We complement this analysis with a closed-form characterization
of the gradient rotation induced by safety projection and establish its monotonicity in the
projection strength and in the angle between reward and cost gradients. Finally, we compare
gradient projection and Lagrangian relaxation across continuous control Safety-Gymnasium
environments empirically.
URL: https://openreview.net/forum?id=EkoNIO1iye
---
Title: Learning without training: The implicit dynamics of in-context learnings
Abstract: One of the most striking features of Large Language Models (LLMs) is their ability to learn in-context. Namely at inference time an LLM is able to learn new patterns without any additional weight update when these patterns are presented in the form of examples in the prompt, even if these patterns were not seen during training. The mechanisms through which this can happen are still largely unknown. In this work, we show that the stacking of a self-attention layer with an MLP allows the transformer block to implicitly modify the weights of the MLP layer according to the context.
We argue through theoretical analysis and experimentation that this simple mechanism may help explain why LLMs demonstrate capabilities of in-context learning, beyond what is captured during training.
Specifically, we show that a standard forward pass with context is mathematically equivalent to a forward pass without context but with the MLP weights updated by a minimal low-rank update representing the context.
URL: https://openreview.net/forum?id=Mx0ZcLznPS
---
Title: ABBEL: Learning Natural-Language Belief States for Memory-Efficient Interaction
Abstract: As the time horizons of sequential decision-making tasks grow, keeping full interaction histories in model context becomes increasingly costly. Recent work reduces context lengths by instead conditioning decision-making agents on recursively updated natural-language summaries, which are concise and interpretable. However, they underperform agents with access to the full context, suggesting that they fail to generate sufficient summaries. To address this we propose ABBEL, a recursive summarization framework that isolates and directly supervises each summary's information contents in the form of explicit natural-language belief states. First, we analyze the belief states generated by frontier models under ABBEL across five domains, and verify that performance is often degraded due to omitting or incorrectly updating information. We also discover settings where models use memory inefficiently by retaining extraneous information. We target these limitations by fine-tuning with two RL-based methods: belief grading, which reduces update errors by rewarding belief generations based on their information content, and peak belief penalties, which encourage compressing the beliefs with the greatest memory footprints. We demonstrate that these methods significantly reduce the performance gap with full context models, and enable ABBEL to outperform prior memory agent work by 40% while using 67% of the memory.
URL: https://openreview.net/forum?id=3j6wBsxlSS
---
Title: Scale-TBPS: Unified Multi-Dataset Training with Noise-Aware Curation and Scalable Discriminative Identity Learning
Abstract: Text-Based Person Search (TBPS) has seen significant progress with vision–language models (VLMs), yet it remains constrained by limited training data and the fact that VLMs are not inherently pre-trained for pedestrian-centric recognition. Existing TBPS methods therefore rely on dataset-centric fine-tuning to handle distribution shift, resulting in multiple independently trained models for different datasets. While synthetic data can increase the scale needed to fine-tune VLMs, it still needs dataset-specific adaptation. This motivates a fundamental question: can we train a single unified TBPS model across multiple datasets? We show that naive joint training over all datasets remains sub-optimal because current training paradigms do not scale to a large number of unique person identities and are vulnerable to noisy image–text pairs. To address these challenges, we propose Scale-TBPS with two contributions: (i) a noise-aware unified dataset curation strategy that cohesively merges diverse TBPS datasets; and (ii) a scalable discriminative identity learning framework that remains effective under a large number of unique identities. Extensive experiments on CUHK-PEDES, ICFG-PEDES, RSTPReid, IIITD-20K, and UFine6926 demonstrate that a single Scale-TBPS model outperforms dataset-centric optimized models and naive joint training. The code is available at: https://anonymous.4open.science/r/Scale-TBPS-52D6/.
URL: https://openreview.net/forum?id=FtkklaKR4K
---
Title: A Causal Perspective on Soft Jump-Diffusion for Time-Series Anomaly Detection
Abstract: Time series anomaly detection is essential for maintaining robustness in dynamic real-world systems. However, most existing methods rely on static distribution assumptions, while overlooking the latent regime-dependent mechanisms and structural shifts that underlie real-world temporal dynamics. This often leads to poor explanation of anomalies and misclassification of environment-induced variations. To address these shortcomings, we propose Causal Soft Jump Diffusion Anomaly Detection (CSJD-AD), a novel framework that models both latent dynamics and soft-gated expected jumps through a structural jump diffusion process. We adopt a causal perspective grounded in environment-conditioned invariance by inferring discrete environment states and condition both the dynamics and jump intensity on them, so the model learns which changes are expected under each regime. By generating paired “expected” (counterfactual) and “observed” (factual) trajectories, the model explicitly contrasts causally consistent behavior with unexplained deviations. Our method achieves state-of-the-art performance across benchmark datasets, demonstrating the importance of incorporating model-implied counterfactual reasoning and jump-aware dynamics into time series anomaly detection.
URL: https://openreview.net/forum?id=QTfZY5G0hz
---
Title: Metric Learning Encoding Models: A Multivariate Framework for Interpreting Neural Representations
Abstract: Understanding how explicit theoretical features are encoded in opaque neural systems is a central challenge now common to neuroscience and AI. We introduce Metric Learning Encoding Models (MLEMs) to address this challenge most directly as a metric learning problem: we fit the distance in the space of theoretical features to match the distance in neural space. Our framework improves on univariate encoding and decoding methods by building on second-order isomorphism methods, such as Representational Similarity Analysis, and extends them by learning a metric that efficiently models feature as well as interactions between them. The effectiveness of MLEM is validated through two sets of simulations. First, MLEMs recover ground-truth importance features in synthetic datasets better than state-of-the-art methods, such as Feature Reweighted RSA (FR-RSA). Second, we deploy MLEMs on real language data, where they show stronger robustness to noise in calculating the importance of linguistic features (gender, tense, etc.). MLEMs are applicable to any domains where theoretical features can be identified, such as language, vision, audition, etc. Code will be made available upon publication.
URL: https://openreview.net/forum?id=mlxGoR0DhU
---
Title: Fast approximate Bayesian multidimensional scaling with consistency guarantees
Abstract: Bayesian multidimensional scaling (BMDS) embeds $n$ objects in a low-dimensional space to approximately preserve an observed dissimilarity matrix. Compared to classic MDS, BMDS is more robust to model misspecification and supports posterior uncertainty quantification and joint estimation within hierarchical models. However, standard BMDS inference is computationally prohibitive, requiring $O(n^2)$ operations per MCMC iteration to evaluate the likelihood. We propose Barnes--Hut BMDS (BH-BMDS), which uses a tree-based approximation to the likelihood and a Gibbs sampler that leverages this structure, remaining compatible with hierarchical extensions. BH-BMDS reduces computational complexity to $O(n \log n)$ while preserving the geometric fidelity of the embedding. We further establish consistency for the stationary measure of BH-BMDS, proving that it concentrates around the true latent configuration even as the total error of the surrogate likelihood diverges. Notably, this consistency holds in the infinite-dimensional limit. We evaluate the approximation on datasets with diverse structure, including air traffic networks, arXiv abstracts, MNIST images and neural activity recordings from mouse models of tau pathology. Across all settings, BH-BMDS closely matches BMDS while achieving substantial computational gains, with approximately 10-fold speedups at $n=1{,}000$ and 70-fold speedups at $n=10{,}000$. These gains increase with $n$, demonstrating strong empirical scalability.
URL: https://openreview.net/forum?id=5ppI8M8rw5
---
Title: From Images to Words: Efficient Cross-Modal Knowledge Distillation to Language Models from Black-box Teachers
Abstract: Knowledge distillation has become the dominant mechanism for transferring capabilities from large models into smaller ones. Yet, despite its widespread success, the transfer remains confined almost entirely within modality boundaries: language models learn from language-only models, vision models learn from vision-only models, and little is known about how knowledge can be distilled across modalities. This limitation is increasingly consequential as modern vision-language models encode rich semantic abstractions about the world that are difficult to acquire from texts alone. Existing cross-modal distillation approaches offer a potential path forward, but typically rely on expensive modality-specific pre-training or multimodal architectures at inference time, limiting their practicality. We introduce ARMADA, a cross-modal distillation framework that transfers knowledge from frozen white-box or black-box multimodal teachers into language-only students. For each text input, ARMADA obtains a frozen teacher representation (e.g., a latent from a text-to-image diffusion model) and employs a lightweight alignment module, TS Aligner, to project teacher and student representations into a shared manifold. The student is trained through a combination of output matching, manifold alignment, and auxiliary supervision objectives, enabling it to internalize cross-modal structure while remaining predictive of downstream tasks. Across twelve natural language understanding tasks, eight reasoning benchmarks, and five instruction-tuning datasets spanning BERT, DeBERTa-v2-1.4B, OPT-1.3B, and LLaMA-{3B,7B,8B} students, ARMADA consistently improves performance, yielding gains of up to 3.4% on NLU tasks and 2.6% on reasoning benchmarks while using only 0.8\% of the trainable parameters required by competing distillation methods. Through controlled comparisons against five degraded-teacher baselines and a capacity-matched control, we show that these gains cannot be explained by additional parameters or generic regularization alone: ARMADA separates significantly ($p<0.1$) from the undistilled baseline and from every fully-degraded teacher, with prediction-level evidence of semantic transfer on grounded inference tasks. Taken together, our results suggest that multimodal foundation models contain transferable semantic structure that can be imported into language-only systems without multimodal pre-training or inference-time multimodal components.
URL: https://openreview.net/forum?id=7DqHteSuqg
---
Title: Steering the Language Axis: From Linear Decodability to Causal Control
Abstract: Despite the impressive multilingual capabilities of Large Language Models, the latent dynamics dictating
language selection remain poorly understood. In this work, we ask whether language identity is merely
linearly decodable from hidden states, or if it can be causally controlled by a compact activation direction.
We conduct an exhaustive causal intervention analysis across multiple model families, including
Qwen 3.5-2B and Llama-3.2-1B-Instruct, isolating PCA-derived "language axes" to perform steering and ablation experiments across 1.26 million generations on the FLORES-200 dataset.
Steering along these geometric directions reliably forces language switching in both cross-script (English to Chinese) and same-script (English to Spanish) settings, whereas equal-magnitude random
perturbations yield virtually no effect. Our layerwise analysis reveals that language commitment
is highly localized and explicitly language-pair-dependent. While English to Chinese switching resists
early intervention and steers easily in the later layers, the English-Spanish transition shifts earlier, displaying a distinct, bimodal sensitivity. Furthermore, targeted ablation uncovers a fundamental reversion to English: once the language signal is removed, the model falls back to English regardless of the input prompt. Ultimately, these findings demonstrate that language decision boundaries function during inference as causally active features that are direction-dependent and layer-specific.
URL: https://openreview.net/forum?id=09BeLFvinO
---
Title: Score-Based Metropolis-Hastings Algorithms
Abstract: In this paper, we introduce a new approach for integrating score-based models with the Metropolis-Hastings algorithm. Score-based diffusion models have emerged as state-of-the-art methods for generative modeling, achieving remarkable performance in accurately learning the score function from data. However, these models lack an explicit energy function, making the Metropolis-Hastings adjustment step inaccessible. Consequently, the unadjusted Langevin algorithm is often used for sampling using estimated score functions.
The lack of an energy function prevents the application of the Metropolis-adjusted Langevin algorithm and other Metropolis-Hastings methods, restricting the use of acceptance-function-based algorithms. We address this limitation by introducing a new loss function based on the detailed balance condition, allowing the estimation of the Metropolis-Hastings acceptance probabilities given a learned score function. We demonstrate the effectiveness of the proposed method for various scenarios, including sampling from score-based diffusion models and heavy-tail distributions.
URL: https://openreview.net/forum?id=YqNNTt871n
---
Title: Planner-Admissible Graph-PDE Value Extensions for Sparse Goal-Conditioned Planning
Abstract: Sparse goal-conditioned planning with few cost-to-go labels can be cast as graph-PDE Dirichlet extension: extend labels on a small boundary Γ_g to unlabelled vertices so that greedy rollouts reach the goal. We argue the key criterion is planner admissibility, not pointwise error. Within the graph p-Laplacian family, harmonic averaging (p = 2) often fails to preserve local greedy orderings, whereas high-p and Absolutely Minimal Lipschitz Extension (AMLE) / Lipschitz-extremal extensions enter a high-success regime. AMLE is the simple, analyzable p = ∞ endpoint of this family, with a local midrange update and a clean comparison / Lipschitz theory.
Our main theoretical statement is a planner-admissibility certificate: under the operational argmin-Q planner, if at every state visited by the surrogate-greedy rollout the local value error stays below half the true action gap, then the rollout reaches the goal (Theorem 3.3). AMLE instantiates the certificate through a comparison-principle fill-distance bound (Corollary 3.5); harmonic can violate it because its neighbour rankings are boundary-label-weighted hitting-probability rankings, not shortest-path rankings (Lemma 3.8, Lemma 3.9). All p-Laplacian endpoints share a structural maximum principle (Lemma 3.6): no strict interior sinks across p ∈ [2, ∞]. An operator-level identity (Proposition 3.10) shows that shortest-path distance is locally AMLE-compatible on the geodesically extendable interior, while harmonic-compatibility requires a non-generic degree-balance condition n_+ = n_- (Corollary 3.11). A refinement-stable seven-node construction is a positive-margin witness for a local harmonic-ordering failure mode that persists under graph subdivision.
On 120 AntMaze layout-derived graph configurations (Fu et al., 2020), aggregate rollout success is 0.584 for p = 2, 0.903 for p = 4, 0.973 for p = 8, 0.982 for a fixed-budget p = 16 solver, and 0.970 for AMLE. Large finite-p rows are not used for exact endpoint ranking because solver certification is incomplete. On the rollout-weighted decision scope, AMLE reduces the low neighbour-Kendall-τ_nbr tail from 0.064 to 0.015 and the true cost-to-go gap of the surrogate-chosen action from 0.049 to 0.006.
URL: https://openreview.net/forum?id=VqgXUZspoV
---
Title: Mind the Ladder: A Benchmark for Level 1--3 Causal Reasoning in World Models
Abstract: World models based on Joint-Embedding Predictive Architecture (JEPA) have recently
demonstrated emergent physical understanding through Violation-of-Expectation (VoE)
paradigms. However, the “surprise” metric used to evaluate these models, a latent prediction
error, conflates statistical novelty with genuine causal reasoning. I argue that no existing
benchmark allows researchers to observe and measure whether a world model has climbed
Pearl’s Ladder of Causality from Level 1 (Association) through Level 2 (Intervention) to
Level 3 (Counterfactuals). In this paper, I introduce Mind the Ladder, a diagnostic
benchmark and accompanying metric suite designed to test causal fidelity in latent world
models. Distinct from prior environment-level causal benchmarks (Ke et al., 2021; Ahmed
et al., 2021), my framework operates directly in the latent space of a trained world model and
is architecture-agnostic. I ground my approach in the LeWorldModel (LeWM) architec-
ture and the stable-worldmodel codebase, proposing three novel metrics (AAP Surprise
Ratio, Structural Invariance, and AAP Consistency Advantage) that operationalize Pearl’s
Abduction-Action-Prediction cycle directly in latent space. I validate my framework on a
“Glitched Hue Two Room” experiment that tests for causal disentanglement between spuri-
ous correlations and true causal mechanisms. My results demonstrate that surprise alone is
not enough: a model can exhibit high VoE surprise for physical violations while still failing
to pass Level 3 counterfactual tests.
URL: https://openreview.net/forum?id=01WKivD0O9
---
Title: Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe
Abstract: On-policy distillation (OPD) has become a core technique in the post-training of large language models, yet its training dynamics remain poorly understood. This paper provides a systematic investigation of OPD dynamics and mechanisms.
We first identify that two conditions govern whether OPD succeeds or fails: (i) the student and teacher should share compatible thinking patterns; and (ii) even with consistent thinking patterns and higher scores, the teacher must offer genuinely new capabilities beyond what the student has seen during training. We validate these findings through weak-to-strong reverse distillation, showing that same-family 1.5B and 7B teachers are distributionally indistinguishable from the student’s perspective. Probing into the token-level mechanism, we show that successful OPD is characterized by progressive alignment on high-probability tokens at student-visited states, a small shared token set that concentrates most of the probability mass (97\%--99\%). We further propose two practical strategies to recover failing OPD: off-policy cold start and teacher-aligned prompt selection. Finally, we show that OPD's apparent free lunch of dense token-level reward comes at a cost, raising the question of whether OPD can scale to long-horizon distillation.
URL: https://openreview.net/forum?id=IiwJEHxlA1
---
Title: A Lifecycle-Perspective Survey of Agent Skills in LLMs: From Construction to Ecosystems
Abstract: Skills---modular, reusable units of agent capability ranging from prompt templates to executable artifacts---have emerged as a central abstraction for LLMs to accumulate, compose, and refine experience. Yet research remains fragmented, with rapid progress scattered across construction, usage, training, safety, and evaluation. We present a unified, lifecycle-oriented survey of skill-based agent systems organized into six interconnected dimensions: Construction and Evolution, Access and Composition, Model Integration via Training, Safety, Evaluation and Feedback, and Skill-Native Systems and Ecosystems. A complementary taxonomy characterizes skills by representation, source, granularity, and role. For each dimension, we surface key challenges and forward-looking directions, including lifecycle management, compositional safety, dynamic utility evaluation, and ecosystem standardization.
URL: https://openreview.net/forum?id=7p1BRbvvJN
---
Title: A Reproducibility Study: Revisiting fairGNN-WOD and Extending to Multiple Demographics
Abstract: Most existing fair GNN methods assume full or partial access to sensitive demographic
attributes during training, which may be unrealistic or infeasible in many real-world settings.
FairGNN-WOD (Wang et al., 2025) is a recently proposed two-stage framework that infers
sensitive demographic attributes and enforces fairness across the inferred groups. In this
paper, we reproduce fairGNN-WOD, provide a more detailed explanation of the framework,
and compare its performance with established baselines. We investigate empirical claims
made by the authors, which indicate that their fairGNN-WOD leads to improvements in
fairness while retaining utility performance. Their framework is tested on the same datasets
used in the original paper, along with an additional dataset. Based on our results, we find
that there is partial support for the claims made by the authors. We further investigate their
model through ablation studies, which give a deeper understanding of the performance of the
demographic inference stage. Finally, we extend the method to also be able to account for
multiple demographic attributes. We test this extension on the same datasets, and achieve a
substantial increase in fairness against the baseline models. The code is publicly available.
URL: https://openreview.net/forum?id=gb8VOMuQAW
---