Daily AI Papers — August 04, 2026

1. Would You Walk to the Car Wash? Revealing the Salience Bias of Large Language Models in Commonsense Reasoning

Authors: Zheng Wu, Chenhao Xue, Shijie Zheng, Yijie Lu, Cheng Yang, Zhuosheng Zhang arXiv: arxiv.org/abs/2607.28478 Summary: LLMs over-prioritize explicit inputs like numbers, causing “Salience Bias” where irrelevant distractors crowd out implicit commonsense prerequisites needed to answer everyday reasoning questions. Testing 12 state-of-the-art LLMs, the authors show this is a suppression failure, not a knowledge gap — a context-free probe recovers over 90% of failures, and lightweight inference-time prompting alone substantially closes the gap. Trending because: One of only two genuinely new papers in today’s HF Daily Papers feed; diagnoses a widely-relevant blind spot across all mainstream LLMs and ships a public benchmark (SaliTrap).

2. SGTP: Sampling-based Game-Theoretic Planning for Real-Time Multi-Vehicle Autonomous Racing

Authors: Zhouheng Li, Fangguo Zhao, Mattia Piccinini, Baha Zarrouki, Yuan Gao, Zitong Shan, Johannes Betz, Chen Lv, Lei Xie arXiv: arxiv.org/abs/2607.25388 Summary: Proposes SGTP, a real-time planner combining game-theoretic reasoning with GPU-accelerated sampling of control sequences to generate diverse, competitive racing behavior among multiple autonomous vehicles. It reports a 95.24% win rate and 99.35% task-completion ratio in interactive races, scaling to 10-agent scenarios with ~0.095s mean solve time, and releases an open-source multi-agent racing benchmark. Trending because: Second of the two genuinely fresh HF Daily Papers entries today; strong quantitative results and an open benchmark release for a niche but active robotics/autonomy subfield.

3. Rethinking Inference-Time Scaling in Local Computer-Use Agents: Failure Modes and Compute Tradeoffs

Authors: Woongkyu Lee, Jungwook Choi arXiv: arxiv.org/abs/2607.28573 Summary: Examines how inference-time compute scaling behaves for computer-use agents run locally, analyzing where scaling breaks down and what compute/accuracy tradeoffs emerge. Directly relevant to the fast-growing computer-use-agent space where cost and latency are practical deployment blockers. Trending because: Computer-use agents are one of the hottest active research threads; failure-mode analyses of inference-time scaling address a widely-shared practical pain point.

4. How Benchmarks Mis-Score Computer-Use Agents

Authors: Zihan Dong, Zhiyuan Ma, Zekun Wang, Yunqing Li, Zirou Liu, Ruixuan Deng, Qishi Zhan, Rui Qian arXiv: arxiv.org/abs/2607.28367 Summary: Argues that current benchmarks systematically mis-score computer-use agents, identifying scoring artifacts that distort comparisons between agent systems. A methodological critique aimed at improving evaluation rigor in a subfield where benchmark trust is currently shaky. Trending because: Companion-topic paper to #3 above — both landed in the same arXiv cs.AI batch, reflecting a coordinated push to scrutinize computer-use agent evaluation.

5. InfoOps Bench: A Live Information Operations Safety Benchmark

Authors: Dorian Quelle, Lisa-Maria Neudert, Jonathan Bright, John Gallacher arXiv: arxiv.org/abs/2607.28503 Summary: Introduces a live benchmark for evaluating AI systems’ vulnerability to and role in information operations (disinformation/influence campaigns), an increasingly urgent AI-safety concern. Framed as a continuously updated (“live”) benchmark rather than a static one-off dataset. Trending because: Addresses AI safety and societal risk, a top-tier significance topic, with a design choice (live benchmark) intended to resist rapid obsolescence.

6. A Foundation Model of Numerical Intelligence with Cross-Disciplinary Generalization

Authors: Chenghan Wu, Zongmin Yu, Liu Yang arXiv: arxiv.org/abs/2607.28432 Summary: Proposes a foundation model specifically targeting numerical/quantitative reasoning that generalizes across disciplines rather than being tuned to one domain. Part of a broader trend toward specialized foundation models for structured reasoning tasks that general LLMs still handle unreliably. Trending because: Numerical reasoning remains a persistent LLM weak spot; a dedicated foundation-model approach with claimed cross-disciplinary transfer is a notable architectural bet.

7. When Derived Measurements Mislead: Quantifying and Mitigating LLM Over-Trust with Privileged-Modality Reliability Evidence

Authors: Zongheng Guo, Tao Chen, Tianli Li, Mingzhe Cui, Yang Jiao, Lei Xie, Yi Pan, Xiao Hu, Manuela Ferrario arXiv: arxiv.org/abs/2607.28421 Summary: Studies how LLMs over-trust derived/computed measurements versus raw evidence, proposing a “privileged-modality reliability” mechanism to quantify and correct this over-trust. Ships code alongside a substantial empirical evaluation (19 tables). Trending because: Extends the growing “LLM calibration/over-trust” literature into a specific, underexplored failure mode (derived vs. raw evidence), with released code lowering the bar for follow-up work.

8. GLM-RAG: Graph Language Models for Graph-Based Retrieval-Augmented Generation

Authors: Maya Arseven, Anette Frank, Beni Egressy, Johann Higl, Moritz Plenz arXiv: arxiv.org/abs/2607.28397 Summary: Combines graph language models with retrieval-augmented generation to reason over graph-structured knowledge rather than flat text passages. Targets the RAG-over-structured-data niche, which is gaining traction as enterprises push RAG beyond plain documents. Trending because: Graph-RAG is a fast-growing applied-AI niche as teams look to ground LLMs in structured/relational data instead of just unstructured text.

9. Beyond Rephrasing: Book-Level Organization Improves Synthetic Textbook Data for Mid-Training

Authors: Jiawen Tao, Miao Peng, Yaoming Li, Xiaokun Yuan, Mengzhou Wu, Wenhan Yu, Guoan Wang, Nuo Chen, Tong Yang, Maxm Pan arXiv: arxiv.org/abs/2607.28109 Summary: Shows that structuring synthetic training data at the “book level” (coherent long-form organization) rather than simple sentence-level rephrasing improves outcomes for LLM mid-training. A 31-page study with substantial empirical backing (11 tables). Trending because: Synthetic data quality for pretraining/mid-training is one of the most consequential open problems in scaling LLMs, and this offers a concrete, testable recipe improvement.

10. Group-Reflective Self-Distillation for Agentic Reinforcement Learning

Authors: Binbin Zheng, Zijun Xie, Guanqun Zhao, Enlei Gong, Xing Ma, Xiaoliang Fu, Zeyu Chen arXiv: arxiv.org/abs/2607.28076 Summary: Introduces a self-distillation method that uses group-level reflection to improve agentic RL training for LLM agents, aiming to make agent policies more sample-efficient and self-correcting. Fits the broader “self-improving agent” research wave. Trending because: Agentic RL and self-distillation are two of the most active current LLM-training research threads; combining them targets a clear efficiency gap.

11. SKILL-KD: Contrastive Skill Distillation for LLM Agents

Authors: Qiming Shi, Yibo Dou, Jiawen Zhu, Yulong Tao, Linbo Jin, Zhaolu Kang, Yunfan Zhou, Di Weng arXiv: arxiv.org/abs/2607.28048 Summary: Proposes contrastive distillation of discrete “skills” from larger to smaller LLM agents, aiming to transfer agentic capability more efficiently than standard knowledge distillation. Targets cheaper deployment of capable agent models. Trending because: Skill-level (vs. token-level) distillation is a natural next step as agentic LLM deployment costs become a bigger practical concern.

12. Lightning OPD 2.0: Mitigating Style Bias in Cross-Teacher On-Policy Distillation for Large Reasoning Models

Authors: Yecheng Wu, Song Han, Han Cai arXiv: arxiv.org/abs/2607.28449 Summary: An update to on-policy distillation for large reasoning models that addresses “style bias” introduced when distilling from a different teacher model, improving cross-teacher transfer fidelity. Comes from a team with a strong track record in efficient-model research (Song Han’s lab). Trending because: On-policy distillation for reasoning models is a hot efficiency technique post-o1/R1-style training; a “2.0” iteration signals active, validated follow-up work.

13. Beyond a Single Judge: Simulating Social Persona Panels for Generative UI Evaluation

Authors: Zheng Wu, Yibo Luo, Pu Zhang, Cheng Yang, Zhuosheng Zhang arXiv: arxiv.org/abs/2607.28439 Summary: Proposes evaluating generative UI outputs using simulated panels of diverse social personas instead of a single LLM judge, aiming to better capture varied human preferences. Shares authors with paper #1 in today’s list, suggesting a coordinated research program on LLM evaluation robustness. Trending because: LLM-as-judge reliability is a widely-debated evaluation bottleneck, and generative UI is a fast-emerging application area needing better assessment methods.

14. HARGO: Heterogeneity-Aware Reward-Guided Optimization for RL Post-Training of LLMs on HPC Tasks

Authors: Tiangang Li, Xiangbo Tian arXiv: arxiv.org/abs/2607.28301 Summary: Proposes a reward-guided RL post-training method that accounts for hardware/task heterogeneity when optimizing LLMs specifically for HPC (high-performance computing) tasks. A domain-specialization play for LLM post-training beyond general chat/coding use cases. Trending because: Extends RL post-training techniques into a less-explored applied domain (HPC), reflecting the broader trend of specializing reasoning-RL recipes per vertical.

15. From Expert Reduction to Behavioral Divergence: Tracing Numerical State through Sparse MoE Inference

Authors: Tianyang Zhu arXiv: arxiv.org/abs/2607.28097 Summary: Investigates how numerical precision/state changes propagate through sparse Mixture-of-Experts inference, tracing how expert routing reduction leads to behavioral divergence in outputs. Relevant to understanding and debugging quantized/compressed MoE deployment. Trending because: MoE architectures now underpin many frontier models, and interpretability/robustness of sparse routing under numerical pressure is an increasingly practical deployment concern.

16. Contrastive Reinforced Policy Optimization via Privileged Self-Distillation

Authors: Xingjian Wu, Junlin Liu, Xingchen Liu, Xuhang Zhu, Jianing Wang, Linsen Guo, Xiaoyu Li, Xuezhi Cao, Xunliang Cai arXiv: arxiv.org/abs/2607.28026 Summary: Combines contrastive learning with reinforced policy optimization and privileged self-distillation to improve RL-based LLM training, likely aimed at more stable/efficient policy gradients. Large industry-style author list (Xunliang Cai is a known Meituan LLM researcher) suggests an applied, production-oriented contribution. Trending because: Another entry in the fast-moving RL-post-training-for-LLMs space, combining three distinct optimization techniques into one recipe.

17. Building a User Foundation Model for the Open Web

Authors: Solal Vernier, Ivan Can Arisoy, Merwan Barlier, Blaž Škrlj arXiv: arxiv.org/abs/2607.28019 Summary: Proposes a foundation model trained to represent user behavior across the open web, aimed at generalizable personalization/recommendation applications. Accepted to RecSys 2026, indicating peer-reviewed validation in the recommender-systems community. Trending because: “Foundation model” framing applied to user/behavioral modeling (rather than language or vision) reflects the broader trend of foundation-model paradigms spreading into recommendation systems; RecSys’26 acceptance adds credibility.

18. TAPO: Transition-Aware Policy Optimization for LLM Agents

Authors: Cong Li, Peixi Peng, Yisen Zhao, Xinyu Hu, Shudong Liu, Zhan Su, Zhuojian Li arXiv: arxiv.org/abs/2607.27973 Summary: Introduces a policy optimization method for LLM agents that explicitly models state transitions rather than treating agent trajectories as flat token sequences. Aimed at improving credit assignment in multi-step agentic tasks. Trending because: Transition/state-aware RL formulations are gaining traction as a fix for poor credit assignment in long-horizon LLM agent training, a widely-recognized limitation of naive RLHF/RLVR setups.

19. Beyond Frame Selection: Generative Latent Evidence Aggregation for Long-Video Understanding

Authors: Bowen Liu, Shuning Wang, Xinpeng Ding, Zhiheng Wu, Bodong Du, Xiaomeng Li arXiv: arxiv.org/abs/2607.28516 Summary: Moves beyond simple frame-selection strategies for long-video understanding by generatively aggregating latent evidence across a video, aiming to better capture information that’s spread across many frames. Targets a persistent bottleneck in long-video multimodal LLMs. Trending because: Long-video understanding is a major open challenge for multimodal LLMs as context windows grow; generative aggregation is a notable departure from the dominant frame-sampling paradigm.

20. Scaling Vision-Language Models Is Not Enough to Mitigate Bias

Authors: Ioannis Sarridis, Ioannis Kompatsiaris, Symeon Papadopoulos arXiv: arxiv.org/abs/2607.28211 Summary: Demonstrates empirically that simply scaling up vision-language models does not reduce (and may not address) bias in their outputs, challenging a common assumption that scale alone improves fairness. A cautionary, evaluation-driven contribution to VLM bias research. Trending because: Directly challenges the “scale solves everything” narrative in a high-stakes area (bias/fairness), relevant to any team relying on larger VLMs as a fairness fix.