🤗 Daily Paper(2025-10-13)

[deep.di...@gmail.com]

🤗 Daily Paper Newsletter Hope you found some gems! This newsletter delivers you the curated list of papers by 🤗 Daily Papers. project page 🤗 daily paper Instant4D: 4D Gaussian Splatting in Minutes Published at 2025-10-01 #ML The authors present Instant4D, a system that efficiently reconstructs scenes from casual video sequences in minutes using native 4D representation and deep visual SLAM, without requiring calibrated cameras or depth sensors. This method significantly reduces model size, speeds up processing, and demonstrates strong performance on various benchmarks, making it applicable to both controlled and real-world video data.... Read More One Patch to Caption Them All: A Unified Zero-Shot Captioning Framework Published at 2025-10-03 #ML The authors propose a new framework for zero-shot captioning that focuses on individual image patches instead of the whole image, allowing for more flexible and detailed image descriptions. They find that using meaningful, dense visual features from backbones like DINO is crucial for achieving top performance in various region-based captioning tasks, and their models outperform other methods in zero-shot dense, region-set, and trace captioning tasks.... Read More Bridging Reasoning to Learning: Unmasking Illusions using Complexity Out of Distribution Generalization Published at 2025-10-06 #ML The authors propose a new framework called Complexity Out of Distribution (Complexity OoD) generalization to define and measure reasoning in AI models. This framework helps distinguish between tasks that can be solved with simple processing and those that require more complex, step-by-step reasoning, and provides recommendations for improving reasoning abilities in AI models.... Read More Progressive Gaussian Transformer with Anisotropy-aware Sampling for Open Vocabulary Occupancy Prediction Published at 2025-10-06 #ML The study presents PG-Occ, a new framework for predicting 3D occupancy in open-vocabulary scenarios, which improves the representation of fine-grained scene details through progressive online densification and an anisotropy-aware sampling strategy, resulting in improved performance and more efficient scene understanding compared to existing methods.... Read More TAG:Tangential Amplifying Guidance for Hallucination-Resistant Diffusion Sampling Published at 2025-10-06 #ML The authors present a new method called Tangential Amplifying Guidance (TAG) that improves image generation by diffusion models without modifying the model's architecture or adding significant computational overhead. TAG works by using an intermediate sample as a reference and adjusting the generation process to reduce inconsistencies and improve sample quality.... Read More A Goal Without a Plan Is Just a Wish: Efficient and Effective Global Planner Training for Long-Horizon Agent Tasks Published at 2025-10-07 #ML The study presents a new method called EAGLET to train a global planner for long-term tasks using large language models, which helps agents avoid trial-and-error and hallucinations. EAGLET uses a two-step process involving high-quality plan synthesis and rule-based reinforcement learning, resulting in improved performance and reduced training costs compared to existing methods.... Read More D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI Published at 2025-10-07 #ML The authors propose a new method called D2E that uses data from desktop environments, specifically gaming, to pretrain models for embodied AI tasks. By collecting and compressing diverse desktop interactions, predicting game events, and transferring desktop-pretrained representations to physical tasks, the proposed framework achieves high success rates in manipulation and navigation benchmarks, demonstrating the potential of desktop pretraining for robotics.... Read More Webscale-RL: Automated Data Pipeline for Scaling RL Data to Pretraining Levels Published at 2025-10-07 #ML The authors present a method to create a large and diverse dataset of question-answer pairs for reinforcement learning (RL) by converting extensive pre-training documents, which is a major challenge in RL. The resulting Webscale-RL dataset, containing 1.2 million examples, enables more efficient RL training, outperforming continual pre-training with up to 100 times fewer tokens.... Read More MONKEY: Masking ON KEY-Value Activation Adapter for Personalization Published at 2025-10-08 #ML The study presents a method to improve personalized diffusion models by using automatic masks to focus on the subject rather than the background, leading to more accurate image generation based on text prompts. This approach outperforms other test time personalization methods in terms of prompt and source image alignment.... Read More Parallel Test-Time Scaling for Latent Reasoning Models Published at 2025-10-08 #ML The study presents two new methods, Monte Carlo Dropout and Additive Gaussian Noise, for sampling in continuous vector spaces used in latent reasoning models, and introduces a Latent Reward Model for aggregating outcomes. Experiments show that these methods improve model performance and offer new directions for scalable inference in continuous spaces.... Read More Temporal Prompting Matters: Rethinking Referring Video Object Segmentation Published at 2025-10-08 #ML This study proposes a new framework called Tenet to improve the efficiency of segmenting objects in videos based on a query sentence. By using pre-trained object detectors and trackers, Tenet generates and selects high-quality temporal prompts to guide image-based segmentation models, resulting in accurate and efficient object segmentation in videos.... Read More ACE: Attribution-Controlled Knowledge Editing for Multi-hop Factual Recall Published at 2025-10-09 #ML The study identifies a limitation in existing knowledge editing methods for large language models, specifically their poor performance in multi-hop factual recall due to an oversight in how chained knowledge is represented and utilized at the neuron level. The researchers propose a new framework called ACE, which uses neuron-level attribution to identify and edit critical pathways, significantly improving multi-hop knowledge editing performance compared to state-of-the-art methods.... Read More ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping Published at 2025-10-09 #ML The authors present ARES, a framework that improves multimodal reasoning models by adjusting their exploration based on task difficulty. ARES uses high window-entropy tokens to identify important moments in reasoning and optimizes exploration through a two-stage training process, resulting in better performance and efficiency compared to existing models.... Read More BEAR: Benchmarking and Enhancing Multimodal Language Models for Atomic Embodied Capabilities Published at 2025-10-09 #ML The authors present BEAR, a comprehensive benchmark that evaluates multimodal language models on various embodied capabilities like perception, spatial reasoning, and planning. They also introduce BEAR-Agent, an improved multimodal agent that enhances the performance of language models on this benchmark, leading to better performance in simulated environments.... Read More Better Together: Leveraging Unpaired Multimodal Data for Stronger Unimodal Models Published at 2025-10-09 #ML This study presents a new method, UML, for training models using unpaired multimodal data, which can improve representation learning in a target modality without needing paired datasets. By processing inputs from different modalities while sharing parameters, the model benefits from cross-modal structure, leading to better performance in downstream tasks for various unimodal targets like images and audio.... Read More BigCodeArena: Unveiling More Reliable Human Preferences in Code Generation via Execution Published at 2025-10-09 #ML The study presents BigCodeArena, a new platform for evaluating code generation models by enabling human interaction with code execution. It collects and analyzes over 14,000 conversations across 10 LLMs and identifies preferences in fine-grained domains. Two benchmarks, BigCodeReward and AutoCodeArena, are introduced to assess LLM performance, with the latter being fully automated.... Read More DISCO: Diversifying Sample Condensation for Efficient Model Evaluation Published at 2025-10-09 #ML The study presents DISCO, a method for selecting the top-k samples that maximize model disagreements, which is simpler and more efficient than traditional clustering-based approaches for evaluating machine learning models. DISCO has shown empirical gains over prior methods and achieved state-of-the-art results in performance prediction across various benchmarks.... Read More Don't Waste Mistakes: Leveraging Negative RL-Groups via Confidence Reweighting Published at 2025-10-09 #ML This study presents a new method called LENS, which improves the use of negative groups in reinforcement learning with verifiable rewards. By assigning confidence-dependent rewards to incorrect responses, LENS makes negative groups informative and boosts the performance of large language models on reasoning tasks.... Read More Formalizing Style in Personal Narratives Published at 2025-10-09 #ML The authors propose a new method to study the unique language styles of personal narratives by combining functional linguistics, computer science, and psychology. They use language models to extract linguistic features and apply this framework to analyze dream narratives, revealing patterns that reflect the author's psychological states.... Read More GTAlign: Game-Theoretic Alignment of LLM Assistants for Mutual Welfare Published at 2025-10-09 #ML The researchers present a new framework called GTAlign that uses game theory to improve the alignment between Large Language Models and users, ensuring that responses are mutually beneficial and socially efficient. This framework enhances reasoning efficiency, answer quality, and user satisfaction compared to existing methods.... Read More LLM4Cell: A Survey of Large Language and Agentic Models for Single-Cell Biology Published at 2025-10-09 #ML A survey of 58 language and agentic models for single-cell biology research is presented, categorized into five families and mapped to eight analytical tasks. The study analyzes benchmark suitability, data diversity, and ethical or scalability constraints, and evaluates models across 10 domain dimensions to provide an integrated view of language-driven single-cell intelligence.... Read More LightReasoner: Can Small Language Models Teach Large Language Models Reasoning? Published at 2025-10-09 #ML The study presents LightReasoner, a new framework that uses smaller language models to teach larger ones, improving reasoning accuracy by up to 28.1% while reducing time, resources, and token usage significantly. This approach makes advancing reasoning in large language models more scalable and efficient.... Read More Pseudo2Real: Task Arithmetic for Pseudo-Label Correction in Automatic Speech Recognition Published at 2025-10-09 #ML The paper presents a method to improve automatic speech recognition in unseen accents and domains without needing target ground truth data. The method uses two ASR models trained on real and pseudo-labeled data to create a correction vector that reduces errors in the pseudo-labeled target model, resulting in a significant improvement in recognition accuracy.... Read More R-Horizon: How Far Can Your Large Reasoning Model Really Go in Breadth and Depth? Published at 2025-10-09 #ML The authors propose a new method called R-HORIZON to test the long-term reasoning abilities of large reasoning models by creating complex tasks with interdependent problems. They find that even advanced models struggle with these tasks, and using R-HORIZON to train models improves their performance on both long-horizon and standard reasoning tasks.... Read More ReviewerToo: Should AI Join The Program Committee? A Look At The Future of Peer Review Published at 2025-10-09 #ML This study presents ReviewerToo, a system that uses AI to help with peer review, aiming to improve consistency and efficiency. By comparing AI-generated reviews to human reviews, the researchers found that AI can be quite accurate and helpful, especially in checking facts and literature, but still needs human expertise for more complex evaluations.... Read More Thinking with Camera: A Unified Multimodal Model for Camera-Centric Understanding and Generation Published at 2025-10-09 #ML The authors present Puffin, a model that combines language and camera perspectives to interpret and generate scenes from various viewpoints. By treating camera as language and using a large-scale dataset, Puffin outperforms specialized models in camera-centric tasks and can generalize to diverse cross-view tasks with instruction tuning.... Read More Understanding DeepResearch via Reports Published at 2025-10-09 #ML The study presents a new method called DeepResearch-ReportEval to evaluate advanced AI systems that conduct research by analyzing their research reports. This approach measures the quality, relevance, and accuracy of the reports and is used to compare four leading AI systems, revealing differences in their performance and design.... Read More Which Heads Matter for Reasoning? RL-Guided KV Cache Compression Published at 2025-10-09 #ML This study proposes a new method called RLKV to identify important heads in large language models that are crucial for reasoning, while compressing the less important ones to reduce cache size. The method uses reinforcement learning to optimize the relationship between each head's cache usage and reasoning quality, resulting in significant cache reduction and near lossless performance compared to uncompressed results.... Read More Adaptive Attacks on Trusted Monitors Subvert AI Control Protocols Published at 2025-10-10 #ML This study reveals a significant vulnerability in AI control protocols, where an untrusted model can evade diverse monitors and complete malicious tasks using a simple adaptive attack vector. The attack works universally against current protocols that rely on a monitor and even backfires against the recent Defer-to-Resample protocol, highlighting the need for improved evaluation of future AI control mechanisms.... Read More AutoPR: Let's Automate Your Academic Promotion! Published at 2025-10-10 #ML The study presents AutoPR, a new task that converts research papers into engaging promotional content, and PRBench, a benchmark for evaluating AutoPR. They also introduce PRAgent, a framework that automates AutoPR, which outperforms LLM pipelines in engagement metrics such as watch time, likes, and overall engagement.... Read More Dyna-Mind: Learning to Simulate from Experience for Better AI Agents Published at 2025-10-10 #ML This study proposes a two-stage training framework, Dyna-Mind, to enhance AI agents' performance in complex interactive environments by teaching them to simulate alternative futures before acting. The framework includes a reasoning method, ReSim, to generate structured reasoning traces and an online reinforcement learning method, Dyna-GRPO, to strengthen the agent's simulation and decision-making ability. Experiments show that Dyna-Mind improves AI agents' reasoning, planning, and action in vari... Read More Hybrid-grained Feature Aggregation with Coarse-to-fine Language Guidance for Self-supervised Monocular Depth Estimation Published at 2025-10-10 #ML The authors present a new method called Hybrid-depth that combines foundation models to improve depth estimation in self-supervised monocular vision. By integrating semantic context from CLIP and spatial details from DINO, their coarse-to-fine approach enhances depth perception, outperforming state-of-the-art methods on the KITTI benchmark.... Read More KORMo: Korean Open Reasoning Model for Everyone Published at 2025-10-10 #ML Researchers created a large, fully open bilingual language model called KORMo-10B specifically for Korean, using mostly synthetic data. They found that synthetic data can support long-term training without issues and allows the model to perform reasoning and follow instructions in Korean as well as existing multilingual models, while also providing all the necessary components for others to replicate and build upon their work.... Read More MRMR: A Realistic and Expert-Level Multidisciplinary Benchmark for Reasoning-Intensive Multimodal Retrieval Published at 2025-10-10 #ML The MRMR benchmark is introduced, which is the first expert-level multidisciplinary multimodal retrieval benchmark that requires intensive reasoning. It has 1,502 queries across 23 domains, with images that need deeper interpretation and a new task called Contradiction Retrieval. The benchmark also features multi-image queries and mixed-modality corpus documents, offering a more realistic setting compared to earlier benchmarks.... Read More Mind-Paced Speaking: A Dual-Brain Approach to Real-Time Reasoning in Spoken Language Models Published at 2025-10-10 #ML The authors propose a new method called Mind-Paced Speaking that allows spoken language models to think and speak simultaneously, improving reasoning performance and reducing latency. This is achieved by dividing the task into two parts: one for high-level reasoning and another for fluent speech generation, similar to how the human brain works.... Read More Mitigating Overthinking through Reasoning Shaping Published at 2025-10-10 #ML This research proposes a new method called Group Relative Segment Penalization (GRSP) to improve the efficiency and accuracy of large reasoning models. GRSP focuses on the granularity of supervision and regulates reasoning at a step-level, leading to better token efficiency and stability in training, especially for more challenging problems, without significantly sacrificing accuracy.... Read More Multimodal Prompt Optimization: Why Not Leverage Multiple Modalities for MLLMs Published at 2025-10-10 #ML This study addresses the limitation of prompt optimization approaches being confined to text, which hinders the full potential of multimodal language models (MLLMs). The researchers propose a unified framework called Multimodal Prompt Optimizer (MPO) that jointly optimizes multimodal prompts and guides prompt selection, outperforming text-only optimization methods across various modalities like images, videos, and molecules.... Read More PhysToolBench: Benchmarking Physical Tool Understanding for MLLMs Published at 2025-10-10 #ML The study presents PhysToolBench, a new benchmark to measure the understanding of physical tools by advanced language models. The benchmark tests the models' abilities in recognizing tools, understanding their functions, and creating new ones, revealing a gap in their knowledge that the researchers aim to address.... Read More SpaceVista: All-Scale Visual Spatial Reasoning from mm to km Published at 2025-10-10 #ML The authors present a solution to improve all-scale spatial reasoning in various scenarios, addressing challenges in dataset curation and scene modeling. They introduce a dataset, SpaceVista-1M, and a model, SpaceVista-7B, which demonstrate strong generalization and competitive performance in extensive evaluations.... Read More Speculative Jacobi-Denoising Decoding for Accelerating Autoregressive Text-to-image Generation Published at 2025-10-10 #ML This study presents a new method called Speculative Jacobi-Denoising Decoding (SJD2) to speed up the text-to-image generation process. SJD2 enables parallel token generation in autoregressive models by incorporating a denoising process, reducing the number of model forward passes while maintaining image quality.... Read More StatEval: A Comprehensive Benchmark for Large Language Models in Statistics Published at 2025-10-10 #ML The authors present StatEval, a new and detailed test for large language models in statistics, which covers basic to advanced topics. They built this benchmark using a method that combines automation and human review, and found that while some models can do well on simpler problems, they struggle with more complex statistical reasoning tasks.... Read More StreamingVLM: Real-Time Understanding for Infinite Video Streams Published at 2025-10-10 #ML The authors present StreamingVLM, a model that efficiently understands infinite video streams in real-time by maintaining a compact cache and using a supervised fine-tuning strategy. This approach allows for stable, real-time performance and even improves general video question-answering abilities without additional training.... Read More TC-LoRA: Temporally Modulated Conditional LoRA for Adaptive Diffusion Control Published at 2025-10-10 #ML A new method called TC-LoRA is proposed to improve controllable diffusion models by dynamically adjusting the model's weights during the generation process, resulting in better adherence to spatial conditions and enhanced generative fidelity compared to traditional methods.... Read More Published at Read More Tags are generated by Google's Gemini Pro API, and the summary and translation are generated by Upstage's SOLAR mini chat model derived from SOLAR-10.7B open LLM. (Experimental) The full paper is translated in korean with enko-t5-small-v0 model developed by Kim Kihyun. Visit Developer's Social Media