Daily TMLR digest for Dec 09, 2025
TMLR
Accepted papers
===============
Title: B-cos LM: Efficiently Transforming Pre-trained Language Models for Improved Explainability
Authors: Yifan Wang, Sukrut Rao, Ji-Ung Lee, Mayank Jobanputra, Vera Demberg
Abstract: Post-hoc explanation methods for black-box models often struggle with faithfulness and human interpretability due to the lack of explainability in current neural architectures. Meanwhile, B-cos networks have been introduced to improve model explainability by proposing an architecture that removes bias terms and promotes input-weight alignment. Although B-cos networks have shown success in building explainable systems, their application has so far been limited to computer vision models and their associated training pipelines. In this work, we introduce B-cos LMs, i.e., B-cos Language Models (LMs) empowered for natural language processing (NLP) tasks. Our approach directly transforms pre-trained language models into B-cos LMs by combining B-cos conversion and task fine-tuning, improving efficiency compared to previous methods. Automatic and human evaluation results demonstrate that B-cos LMs produce more faithful and human interpretable explanations than post-hoc methods, while maintaining task performance comparable to conventional fine-tuning. Our in-depth analysis explores how B-cos LMs differ from conventionally fine-tuned models in their learning processes and explanation patterns. Finally, we present a first exploration of transforming decoder-only models to B-cos LMs for generation tasks. Our code is available at https://github.com/Ewanwong/bcos_lm.
URL: https://openreview.net/forum?id=c180UH8Dg8
---
Title: How to Upscale Neural Networks with Scaling Law?
Authors: Ayan Sengupta, Yash Goel, Tanmoy Chakraborty
Abstract: Neural scaling laws have revolutionized the design and optimization of large-scale AI models by revealing predictable relationships between model size, dataset volume, and computational resources. Early research established power-law relationships in model performance, leading to compute-optimal scaling strategies. However, recent studies highlighted their limitations across architectures, modalities, and deployment contexts. Sparse models, mixture-of-experts, retrieval-augmented learning, and multimodal models often deviate from traditional scaling patterns. Moreover, scaling behaviors vary across domains such as vision, reinforcement learning, and fine-tuning, underscoring the need for more nuanced approaches. In this survey, we synthesize insights from current studies, examining the theoretical foundations, empirical findings, and practical implications of scaling laws. We also explore key challenges, including data efficiency, inference scaling, and architecture-specific constraints, advocating for adaptive scaling strategies tailored to real-world applications. We suggest that while scaling laws provide a useful guide, they do not always generalize across all architectures and training strategies.
URL: https://openreview.net/forum?id=AL7N0UOfgI
---
Title: COMPASS: COntinual Multilingual PEFT with Adaptive Semantic Sampling
Authors: Noah Flynn
Abstract: Large language models (LLMs) often exhibit performance disparities across languages, with naive multilingual fine-tuning frequently degrading performance due to negative cross-lingual interference. To address this, we introduce COMPASS (COntinual Multilingual PEFT with Adaptive Semantic Sampling), a novel data-centric framework for adapting LLMs to target languages. COMPASS leverages parameter-efficient fine-tuning (PEFT) by training lightweight, language-specific adapters on a judiciously selected subset of auxiliary multilingual data. The core of our method is a distribution-aware sampling strategy that uses multilingual embeddings and clustering to identify semantic gaps between existing training data and a target usage distribution. By prioritizing auxiliary data from under-represented semantic clusters, COMPASS maximizes positive cross-lingual transfer while minimizing interference. We extend this into a continual learning framework, COMPASS-ECDA, which monitors for data distribution shifts in production and dynamically updates adapters to prevent model staleness, balancing adaptation to new data with the preservation of existing knowledge. Across three different model architectures (Phi-4-Mini, Llama-3.1-8B, and Qwen2.5-7B) and multiple challenging multilingual benchmarks (Global-MMLU, MMLU-ProX), including unseen long-context tasks (OneRuler), we demonstrate that COMPASS consistently outperforms baseline methods guided by linguistic similarity, providing an effective, efficient, and sustainable solution for developing and maintaining high-performing multilingual models in dynamic environments.
URL: https://openreview.net/forum?id=oapsbIO1Bd
---
Title: Denoising Pretrained Black-box Models via Amplitude-Guided Phase Realignment
Authors: Hongliang Ni, Tong Chen, Shazia Sadiq, Gianluca Demartini
Abstract: Pre-trained models tend to inherit noisy label information from their training datasets, internalising it as biased knowledge. While learning with label noise has been explored, existing approaches rarely address the mitigation of biased knowledge embedded in pre-trained representations introduced by noisy labels. Moreover, existing denoising methods invariably rely on modifying training datasets or models to improve downstream task performance. However, we observe a growing trend in which both pre-trained models and their training datasets are scaling up significantly and becoming increasingly inaccessible, making modifications ever more infeasible. In this paper, we propose a black-box biased knowledge mitigation method called ``Lorem'', which leverages feature frequency amplitudes to guide phase correction on pre-trained representations, without access to training data or model parameters. We first present empirical evidence that, across different noise levels, the phase components of pre-trained representations are more sensitive to noisy labels than the amplitude components, while discriminative information for classification is primarily encoded in the amplitude. Moreover, we find that the impact of noisy labels on amplitude is global, leading to a gradual loss of discriminative information. Therefore, corrective strategies must be adaptive across the entire frequency spectrum rather than limited to the high-frequency components. Inspired by this observation, we design a method that leverages the amplitude residual to realign phase, thereby removing biased knowledge from pre-trained representations. Experiments on a variety of popular pre-trained vision and language models suggest that, even with a simple linear classifier, our method can enhance downstream performance across a range of in-domain and out-of-domain tasks.
URL: https://openreview.net/forum?id=526fwttJiK
---
Title: FedDUAL: A Dual-Strategy with Adaptive Loss and Dynamic Aggregation for Mitigating Data Heterogeneity in Federated Learning
Authors: Pranab Sahoo, Ashutosh Tripathi, Sriparna Saha, Samrat Mondal
Abstract: Federated Learning (FL) marks a transformative approach to distributed model training by combining locally optimized models from various clients into a unified global model. While FL preserves data privacy by eliminating centralized storage, it encounters significant challenges such as performance degradation, slower convergence, and reduced robustness of the global model due to the heterogeneity in client data distributions. Among the various forms of data heterogeneity, label skew emerges as a particularly formidable and prevalent issue, especially in domains such as image classification. To address these challenges, we begin with comprehensive experiments to pinpoint the underlying issues in the FL training process, such as gradient instability and the emergence of sharp minima in the global model, both of which contribute to performance inconsistencies. Based on our findings, we introduce an innovative dual-strategy approach designed to effectively resolve these issues. First, we introduce an adaptive loss function for client-side training, meticulously crafted to preserve previously acquired knowledge while maintaining an optimal equilibrium between local optimization and global model coherence. Secondly, we develop a dynamic aggregation strategy for aggregating client models at the server. This approach adapts to each client's unique learning patterns, effectively addressing the challenges of diverse data across the network. Our comprehensive evaluation, conducted across three diverse real-world datasets, coupled with theoretical convergence guarantees, demonstrates the superior efficacy of our method compared to several established state-of-the-art approaches.
URL: https://openreview.net/forum?id=8M3XfmNhTZ
---
Title: VirDA: Reusing Backbone for Unsupervised Domain Adaptation with Visual Reprogramming
Authors: Duc-Duy Nguyen, Dat Nguyen
Abstract: Image classification is among the pillars of computer-vision pipelines. While state-of-the-art models excel within their training domains, their performance often deteriorates when transferred to a new, unlabeled setting. Unsupervised domain adaptation (UDA) addresses this challenge by repurposing a well-trained source classifier for the target domain, enabling strong downstream results without the need for additional labeled data. Existing UDA pipelines fine-tune already well-trained backbone parameters for every new source-and-target pair, resulting in the number of training parameters and storage memory growing linearly with each new pair, and also preventing the reuse of these well-trained backbone parameters.
Inspired by recent implications that existing backbones have textural biases, we propose making use of domain-specific textural bias for domain adaptation via visual reprogramming, namely VirDA.
Instead of fine-tuning the full backbone, VirDA prepends a domain-specific visual reprogramming layer to the backbone. This layer produces visual prompts that act as an added textural bias to the input image, adapting its ``style'' to a target domain. To optimize these visual reprogramming layers, we use multiple objective functions that optimize the intra- and inter-domain distribution differences when domain-adapting visual prompts are applied. This process does not require modifying the backbone parameters, allowing the same backbone to be reused across different domains.
We evaluate VirDA on Office-31 and obtain 92.8% mean accuracy with only 1.5M trainable parameters. VirDA surpasses PDA, the state-of-the-art parameter-efficient UDA baseline, by +1.6% accuracy while using just 46% of its parameters. Compared with full-backbone fine-tuning, VirDA outperforms CDTrans and FixBi by +0.2% and +1.4%, respectively, while requiring only 1.7% and 2.8% of their trainable parameters. Relative to the strongest current methods (PMTrans and TVT), VirDA uses ~1.7% of their parameters and trades off only 2.2% and 1.1% accuracy, respectively.
URL: https://openreview.net/forum?id=Qh7or7JRFI
---
New submissions
===============
Title: Socrates Loss: Unifying Confidence Calibration and Classification by Leveraging the Unknown
Abstract: Deep neural networks, despite their high accuracy, often exhibit poor confidence calibration, limiting their reliability in high-stakes applications. Current ad-hoc confidence calibration methods attempt to fix this during training but face a fundamental trade-off: two-phase training methods achieve strong classification performance at the cost of training instability and poorer confidence calibration, while single-loss methods are stable but underperform in classification. This paper resolves this stability-performance trade-off. We propose Socrates Loss, a novel, unified loss function that explicitly leverages uncertainty by incorporating an auxiliary unknown class, whose predictions directly influence the loss function and a dynamic uncertainty penalty. This unified objective allows the model to be optimized for both classification and confidence calibration simultaneously, without the instability of complex, scheduled losses. We provide theoretical guarantees that our method regularizes the model to prevent miscalibration and overfitting. Across four benchmark datasets and multiple architectures, our comprehensive experiments demonstrate that Socrates Loss is not only more stable but also achieves a state-of-the-art balance of accuracy and calibration, often converging faster than existing methods.
URL: https://openreview.net/forum?id=DONqw1KhHq
---
Title: Experience Replay with Random Reshuffling
Abstract: Experience replay is a key component in reinforcement learning for stabilizing learning and improving sample efficiency. Its typical implementation samples transitions with replacement from a replay buffer. In contrast, in supervised learning with a fixed dataset, it is a common practice to shuffle the dataset every epoch and consume data sequentially, which is called random reshuffling (RR). RR enjoys theoretically better convergence properties and has been shown to outperform with-replacement sampling empirically. To leverage the benefits of RR in reinforcement learning, we propose sampling methods that extend RR to experience replay, both in uniform and prioritized settings, and analyze their properties via theoretical analysis and simulations. We evaluate our sampling methods on Atari benchmarks, demonstrating their effectiveness in deep reinforcement learning.
URL: https://openreview.net/forum?id=56cbwigQSj
---
Title: When unlearning is free: leveraging low influence points to reduce computational costs
Abstract: As concerns around data privacy in machine learning grow, the ability to unlearn-or remove- specific data points from trained models becomes increasingly important. While state-of-the-art unlearning methods have emerged in response, they typically treat all points in the forget set equally. In this work, we challenge this approach by asking: do points that have a negligible impact on the model's learning need to be removed? Through a comparative analysis of influence functions across language and vision tasks, we identify subsets of training data with negligible impact on model outputs. Leveraging this insight, we propose an efficient unlearning framework that reduces the size of datasets before unlearning—leading to significant computational savings (up to ~50%) on real-world empirical examples.
URL: https://openreview.net/forum?id=arA4b7O3Wn
---
Title: Improving Visual Discriminability of CLIP for Training-Free Open-Vocabulary Semantic Segmentation
Abstract: Extending CLIP models to semantic segmentation remains challenging due to the misalignment between their image-level pre-training objectives and the pixel-level visual understanding required for dense prediction. While prior efforts have achieved encouraging results by reorganizing the final layer and features, they often inherit the global alignment bias of preceding layers, leading to suboptimal segmentation performance. In this work, we propose LHT-CLIP, a novel training-free framework that systematically exploits the visual discriminability of CLIP across \emph{layer}, \emph{head}, and \emph{token} levels. Through comprehensive analysis, we reveal three key insights: (i) the final layers primarily strengthen image–text alignment with sacrifice of visual discriminability (e.g., last 3 layers in ViT-B/16 and 8 layers in ViT-L/14), partly due to the emergence of anomalous tokens; (ii) a subset of attention heads (e.g., 10 out of 144 in ViT-B/16) display consistently strong visual discriminability across datasets; (iii) abnormal tokens display sparse and consistent activation pattern compared to normal tokens. Based on these findings, we propose three complementary techniques: semantic-spatial reweighting, selective head enhancement, and abnormal token replacement to effectively restore visual discriminability and improve segmentation performance without any additional training, auxiliary pre-trained networks, or extensive hyperparameter tuning. Comprehensive experiments on eight widely used semantic segmentation benchmarks demonstrate that LHT-CLIP achieves substantial performance improvements across diverse scenarios, underscoring its effectiveness and practicality for real-world deployment.
URL: https://openreview.net/forum?id=9spNW3DXg5
---
Title: On the Impact of Hyper-Parameters on the Privacy of Deep Neural Networks
Abstract: The deployment of deep neural networks (DNNs) in many real-world applications leads to the processing of huge amounts of potentially sensitive data. This raises important new concerns, in particular with regards to the privacy of individuals whose data is used by these DNNs. In this work, we focus on DNNs trained to identify biometric markers from images, e.g., gender classification, which have been shown to leak unrelated private attributes at inference time, e.g., ethnicity, also referred to as unintentional feature leakage. Existing literature has tackled this problem through architecture specific and complex techniques that are hard to put into place in practice. In contrast we focus on a very generalizable aspect of DNNs, the hyper-parameters used to train them, and study how they impact the privacy risk. Specifically, we follow a multi-fidelity and multi-objective HPO approach to (i) conduct the first study of the impact of hyper-parameters on the risk of unintended feature leakage (privacy risk); (ii) demonstrate that, for a specific main task, HPO successfully identifies hyper-parameter configurations that considerably reduce the privacy risk at a very low impact on utility, achieving similar result as state-of-the-art techniques only by changing hyper-parameters; and (iii) evidence that there exist hyper-parameter configurations that have a significant impact on the privacy risk, regardless of the choice of main and private tasks, i.e., hyper-parameters that generally better preserve privacy.
URL: https://openreview.net/forum?id=xCCObBkPp9
---
Title: Wasserstein Bounds for generative diffusion models with Gaussian tail targets
Abstract: We present an estimate of the Wasserstein distance between the data distribution and the generation of score-based generative models. The sampling complexity with respect to dimension is $\mathcal{O}(\sqrt{d})$, with a logarithmic constant. In the analysis, we assume a Gaussian-type tail behavior of the data distribution and an $\epsilon$-accurate approximation of the score. Such a Gaussian tail assumption is general, as it accommodates practical target distributions derived from early stopping techniques with bounded support.
The crux of the analysis lies in the global Lipschitz bound of the score, which is shown from the Gaussian tail assumption by a dimension-independent estimate of the heat kernel. Consequently, our complexity bound scales linearly (up to a logarithmic constant) with the square root of the trace of the covariance operator, which relates to the invariant distribution of the forward process.
URL: https://openreview.net/forum?id=QbQ4DtP5vS
---
Title: DiffProbe: Towards a Universal and Cross-Modality Adversarial Robustness Quantification Framework for Black-box Classifiers using Diffusion Models
Abstract: Neural network classifiers have become popular, fundamentally transforming tasks across multiple data modalities such as image processing, natural language understanding, audio recognition, and more. Despite their widespread adoption, a critical challenge that persists is ensuring their robustness against adversarial attacks, which aim to deceive models through subtly modified inputs. This issue is particularly acute when considering interactions across different modalities, a facet that most current studies neglect. Addressing this gap, our paper introduces \textbf{DiffProbe}, the first unified black-box framework for adversarial robustness quantification using synthetically generated data from domain-specific diffusion models. \textbf{DiffProbe} stands out by seamlessly integrating off-the-shelf diffusion models to create a versatile and comprehensive framework tool suitable for a wide range of data types and adversarial scenarios. It is particularly designed for computational efficiency, making it ideal for evaluating black-box models and facilitating remote auditing with minimal requirement—only needing predictions from models on synthetic data. The robustness evaluation of \textbf{DiffProbe} is both theoretically sound and empirically robust, showing high consistency with real-world adversarial attack methods. We have extensively tested \textbf{DiffProbe} across various state-of-the-art classifiers and black-box APIs in domains including facial recognition, text, audio, video, and point cloud data. The results underscore its effectiveness in providing realistic and actionable insights into the adversarial robustness of systems, thus enhancing our understanding of adversarial vulnerabilities and aiding in the development of more secure AI systems across different modalities.
URL: https://openreview.net/forum?id=DlcjsaG7rX
---
Title: Continual Robot Learning via Language-Guided Skill Acquisition
Abstract: To support daily human tasks, robots need to tackle complex, long-horizon tasks and continuously acquire new skills to handle new problems. Deep Reinforcement Learning (DRL) offers potential for learning fine-grained skills but relies heavily on human-defined rewards and faces challenges with long-horizon goals. Task and Motion Planning (TAMP) are adept at handling long-horizon tasks but often need tailored domain-specific skills, resulting in practical limitations and inefficiencies. To overcome these complementary limitations, we propose LG-SAIL (Language Models Guided Sequential, Adaptive, and Incremental Skill Learning), a framework that leverages Large Language Models (LLMs) to synergistically integrate TAMP and DRL for continuous skill learning in long-horizon tasks. Our framework achieves automatic task decomposition, operator creation, and dense reward generation for efficiently acquiring the desired skills. To facilitate new skill learning, our framework maintains a symbolic skill library and utilizes the existing model from semantic-related skills to warm start the training. LG-SAIL demonstrates superior performance compared to baselines across six challenging simulated task domains across two benchmarks. Furthermore, we demonstrate the ability to reuse learned skills to expedite learning in new task domains, and deploy the system on a physical robot platform. More results on website: https://sites.google.com/view/continuallearning.
URL: https://openreview.net/forum?id=oYRNxxGN9u
---
Title: SNAP: Stopping Catastrophic Forgetting in Hebbian Learning with Sigmoidal Neuronal Adaptive Plasticity
Abstract: Artificial Neural Networks (ANNs) suffer from catastrophic forgetting, where learning on new tasks causes the degradation of performance on previous ones. Existing algorithms typically use linear weight updates, where the magnitude of the update is independent of the current weight strength. This contrasts with biological neurons, which at intermediate strengths are very plastic, but consolidate with Long-Term Potentiation (LTP) once they reach a certain strength. We hypothesize that this biological mechanism could mitigate catastrophic forgetting in ANNs. We introduce Sigmoidal Neuronal Adaptive Plasticity (SNAP), an artificial approximation to Long-Term Potentiation for ANNs by having the weights follow a sigmoidal growth behavior, allowing the weights to consolidate and stabilize when they reach sufficiently extreme values. We compare SNAP to linear and exponential weight growth and see that SNAP prevents the forgetting of previous tasks for Hebbian Learning but not for Stochastic Gradient Descent (SGD) based learning.
URL: https://openreview.net/forum?id=Vo0XJGyFQb
---
Title: Learning to Guide Human Decision Makers with Vision-Language Models
Abstract: There is growing interest in AI systems that support human decision-making in high-stakes domains (e.g., medical diagnosis) to improve decision quality and reduce cognitive load. Mainstream approaches pair human experts with a machine-learning model, offloading low-risk decisions to the model so that experts can focus on cases that require their judgment.
This $\textbf{\textit{separation of responsibilities}}$ setup, however, is inadequate for high-stakes scenarios. The expert may end up over-relying on the machine's decisions due to $\textit{anchoring bias}$, thus losing the human oversight that is increasingly being required by regulatory agencies to ensure trustworthy AI. On the other hand, the expert is left entirely unassisted on the (typically hardest) decisions on which the model abstained.
As a remedy, we introduce $\textbf{\textit{learning to guide}}$ (LTG), an alternative framework in which -- rather than taking control from the human expert -- the machine provides $\textit{guidance}$ useful for decision making, and the human is entirely responsible for coming up with a decision.
In order to ensure guidance is $\textit{interpretable}$ and $\textit{task-specific}$, we develop Slog, an approach for turning $\textit{any}$ vision-language model into a capable generator of textual guidance by leveraging a modicum of human feedback.
Our empirical evaluation highlights the promise of Slog on both on a synthetic dataset and a challenging, real-world medical diagnosis task.
URL: https://openreview.net/forum?id=TpqVKoEn9M
---
Title: Not All Rollouts are Useful: Down-Sampling Rollouts in LLM Reinforcement Learning
Abstract: Reinforcement learning with verifiable rewards (RLVR) has emerged as the leading approach for enhancing reasoning capabilities in large language models. However, it faces a fundamental compute and memory asymmetry: rollout generation is embarrassingly parallel and memory-light, whereas policy updates are communication-heavy and memory-intensive. To address this, we introduce **PODS** (**P**olicy **O**ptimization with **D**own-**S**ampling), which decouples rollout generation from policy updates by training only on a strategically selected subset of rollouts, maintaining learning quality while dramatically reducing update costs. We propose a principled subset selection criterion—*max-variance down-sampling*—that maximizes the variance of reward in the selected subset, and provide an efficient $O(n\log n)$ implementation of this rule. Empirically, Group Relative Policy Optimization (GRPO) with PODS achieves the peak test accuracy of vanilla GRPO at least $\mathbf{1.7\times}$ **faster** across the different reasoning benchmarks and hardware configurations we tested.
URL: https://openreview.net/forum?id=MfHOmgqVXM
---
Title: Gradient-Based Program Repair: Fixing Bugs in Continuous Program Spaces
Abstract: Automatic program repair seeks to generate correct code from buggy programs, with most approaches searching the correct program in a discrete, symbolic space of source code tokens.
This symbolic search is fundamentally limited by its inability to directly reason about program behavior.
We introduce Gradient-Based Program Repair (GBPR), a new paradigm that reframes program repair as continuous optimization in a differentiable numerical program space.
Our core insight is to compile symbolic programs into differentiable numerical representations, enabling search in the numerical program space directly guided by program behavior.
To evaluate GBPR, we present RaspBugs, a new benchmark of 1,466 buggy symbolic RASP programs and their respective numerical representations.
Our experiments demonstrate that GBPR can effectively repair buggy symbolic programs by gradient-based optimization in the numerical program space, with convincing repair trajectories.
To our knowledge, we are the first to state program repair as continuous optimization in a numerical program space.
Our work establishes a new direction for program repair research, bridging two rich worlds: continuous optimization and program behavior.
URL: https://openreview.net/forum?id=b0611Iy0QM
---
Title: Hyperspherical Simplex Representations from Softmax Outputs and Logits are Inherently Backward-Compatible
Abstract: Training modern AI models has become increasingly expensive, and updating deployed models can significantly alter the behavior of applications built upon them, due to changes in internal feature representations. In retrieval systems, such updates often require re-indexing the gallery-set by extracting feature vectors for all gallery data, a process that is computationally expensive and time-consuming, especially for large-scale datasets. Existing backward-compatibility methods allow direct comparison between the representations of updated and old models, avoiding the re-indexing of the gallery. However, they typically introduce a dependency on the old model by using auxiliary losses, mapping functions, or specific modifications to the model architecture. In this paper, we show that independently trained models are inherently backward-compatible when hyperspherical simplex representations derived from their softmax outputs or logits are used. Leveraging the geometric structure induced by the softmax function on these features, we define a deterministic projection that preserves their alignment across model updates. We demonstrate that these representations satisfy in expectation the formal definition of backward-compatibility. Without relying on regularization-based training, mapping functions, or modifications to the model architecture, we achieve competitive results on standard compatibility benchmarks involving model updates with new training classes and/or advanced model architectures.
URL: https://openreview.net/forum?id=0eeXx7QZO7
---
Title: Explaining Graph Neural Networks for Node Similarity on Graphs
Abstract: Similarity search is a fundamental task for exploiting information in various applications dealing with graph data, such as citation networks or knowledge graphs.
Prior work on the explainability of graph neural networks (GNNs) has focused on supervised tasks, such as node classification and link prediction. However, the challenge of explaining similarities between node embeddings has been left unaddressed.
We take a step towards filling this gap by formulating the problem, identifying desirable properties of explanations of similarity, and proposing intervention-based metrics that qualitatively assess them.
Using our framework, we evaluate the performance of representative methods for explaining GNNs, based on the concepts of mutual information (MI) and gradient-based (GB) explanations. We find that unlike MI explanations,
GB explanations have three desirable properties. First, they are *actionable*: selecting particular inputs results in predictable changes in similarity scores of corresponding nodes. Second, they are *consistent*: the effect of selecting certain inputs hardly overlaps with the effect of discarding them. Third, they can be pruned significantly to obtain *sparse* explanations that retain the effect on similarity scores. These important findings highlight the utility of our metrics as a framework for evaluating the quality of explanations of node similarities in GNNs.
URL: https://openreview.net/forum?id=zDEwl4zidP
---
Title: Provably Efficient Off-Policy Adversarial Imitation Learning with Convergence Guarantees
Abstract: Adversarial Imitation Learning (AIL) faces challenges with sample inefficiency because of its reliance on sufficient on-policy data to evaluate the performance of the current policy during reward function updates. In this work, we study the convergence properties and sample complexity of off-policy AIL algorithms. We show that, even in the absence of importance sampling correction, reusing samples generated by the $o(\sqrt{K})$ most recent policies, where $K$ is the number of iterations of policy updates and reward updates, does not undermine the convergence guarantees of this class of algorithms. Furthermore, our results indicate that the distribution shift error induced by off-policy updates is dominated by the benefits of having more data available. This result provides theoretical support for the sample efficiency of off-policy AIL algorithms that has been observed in practice.
URL: https://openreview.net/forum?id=OahvMeRgKP
---
Title: Covariance Density Neural Networks
Abstract: Graph neural networks have re-defined how we model and predict on network data but
there lacks a consensus on choosing the correct underlying graph structure on which to
model signals. CoVariance Neural Networks (VNN) address this issue by using the sample
covariance matrix as a Graph Shift Operator (GSO). Here, we improve on the performance
of VNNs by constructing a Density Matrix where we consider the sample Covariance matrix
as a quasi-Hamiltonian of the system in the space of random variables. Crucially, using this
density matrix as the GSO allows components of the data to be extracted at different scales,
allowing enhanced discriminability and performance. We show that this approach allows
explicit control of the stability-discriminability trade-off of the network, provides enhanced
robustness to noise compared to VNNs, and outperforms them in useful real-life applications
where the underlying covariance matrix is informative. In particular, we show that our
model can achieve strong performance in subject-independent Brain Computer Interface
EEG motor imagery classification, outperforming EEGnet while being faster. This shows
how covariance density neural networks provide a basis for the notoriously difficult task of
transferability of BCIs when evaluated on unseen individuals.
URL: https://openreview.net/forum?id=TwCkGi5XFB
---
Title: Aggregation-Free Heterogeneous Federated Learning with Data-Free Knowledge Exchange
Abstract: Heterogeneous Federated Learning (HFL) is a decentralized machine learning paradigm that enables participants to leverage distributed knowledge from diversified environments while safeguarding individual privacy. Recent works that address both data and model heterogeneity still require aggregating model parameters, which restricts architectural flexibility. Knowledge Distillation (KD) has been adopted in HFL to circumvent direct model aggregation by aggregating knowledge, but it depends on a public dataset and may incur information loss when redistributing knowledge from the global model. We propose Federated Knowledge Exchange (FKE), an aggregation-free FL paradigm in which each client acts as both teacher and student, exchanging knowledge directly with peers and removing the need for a global model. To remove reliance on public data, we attach a lightweight embedding decoder that produces transfer data, forming the Data-Free Federated Knowledge Exchange (DFFKE) framework. Extensive experiments show that DFFKE surpasses nine state-of-the-art HFL baselines by up to 18.14%. Code is available in the supplementary material. Anonymous Repo: https://anonymous.4open.science/r/DFFKE-0E0B.
URL: https://openreview.net/forum?id=BLQ9JulhQm
---
Title: PhasorTransformer: Integrating Rotational Inductive Biases for Complex-Valued Sequence Modeling
Abstract: Deep neural networks typically process complex-valued signals—such as RF waveforms or
MRI data—via a convenient approximation: they split the real and imaginary parts into
separate, independent channels. This works, but it ignores the underlying mathematics.
By treating these components as disjoint, standard architectures become blind to the sig-
nal’s algebraic structure, specifically the rotational geometry of the phase. We introduce
the PhasorTransformer to correct this misalignment. Instead of avoiding complex arith-
metic, our architecture embeds it directly into the attention mechanism. We generalize
Rotary Positional Embeddings (RoPE) to the complex plane and apply a Hermitian inner
product to derive a strictly equivariant attention layer; this allows the network to handle
phase shifts naturally rather than relearning them as separate features. On the Long-Range
Arena (Sequential CIFAR-10) and Radio Modulation Classification benchmarks, our ap-
proach matches or outperforms state-of-the-art real-valued baselines. Crucially, it achieves
these results with up to a 20×reduction in parameters, demonstrating that respecting the
holomorphic structure of physical signals provides a massive efficiency advantage.
URL: https://openreview.net/forum?id=u2BDXTmtds
---
Title: Pass@k Metric for RLVR: A Diagnostic Tool of Exploration, But Not an Objective
Abstract: The ability of Large Language Models (LLMs) to perform complex, multi-step reasoning is a central focus of modern AI research. To evaluate and enhance this capability, the pass@k metric, which measures the probability of obtaining at least one correct solution in k independent samples, has received significant attention. Its intuitive appeal has led to its adoption not only as an evaluation standard but also as a direct optimization objective in reinforcement learning. In this paper, we analyze the pass@k objective, derive its gradient, and demonstrate that it is fundamentally a per-example positive reweighting of the simpler pass@1 objective. Our analysis reveals that the pass@k objective provides a vanishing learning signal in situations where exploration is most critical. We further analyze the dynamics of ``exploration collapse'', showing that as the policy concentrates probability mass, the gap between pass@k and pass@1 diminishes. We conclude that while pass@k is a useful diagnostic tool, it may be an unsuitable direct objective for optimization. Instead, mechanisms explicitly encouraging efficient exploration could offer a more effective path forward for reinforcement learning in reasoning tasks.
URL: https://openreview.net/forum?id=TbUWwuvY2b
---