Daily AI Papers — August 02, 2026
Published:
1. OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models
Authors: Qiushi Sun, Kanzhi Cheng, Yian Wang, Bowen Yang, Hang Yan, Liheng Chen, Fangzhi Xu, Zichen Ding, Nuo Chen, et al.
Summary: Introduces a rigorously human-labeled benchmark showing that even frontier VLM judges of computer-use-agent trajectories share a systematic leniency bias, mislabeling failures as successes. The authors release OS-Shepherd-100K and train open 9B/35B reward models that match commercial judges at 30-60% lower cost.
arXiv: arxiv.org/abs/2607.28609 Sources: arXiv (cross-listed cs.AI, cs.CL, cs.CV) Why trending: Triple cross-listed across AI/CL/CV categories — addresses a foundational reliability gap (agent reward-model trust) as computer-use agents scale.
2. Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost, from 1.5B to 7B
Authors: Iliya Mirzaei
Summary: A carefully controlled study (7 methods, 3 model sizes, 2 math benchmarks, bootstrap-tested) finds that self-critique methods like Self-Refine and Reflexion never reliably beat simple repeated sampling once token budgets are matched — and self-inspection methods are often reliably worse.
arXiv: arxiv.org/abs/2607.28576 Sources: arXiv (cross-listed cs.AI, cs.CL, cs.LG) Why trending: Triple cross-listed; a rigorous negative result challenging popular “self-reflection” inference-time scaling techniques.
3. Baikal: Structured Search for Deep Research over Data Lakes
Authors: Dhruv Agarwal, Rishitha Guttapalle Mohan, Aarti Kumari, Ashi Sinha, Athulya Anil, Kavitha Srinivas, Horst Samulowitz, Andrew McCallum
Summary: Casts deep-research agents over heterogeneous data lakes as a budgeted search problem, clustering evidence into semantic regions and adaptively balancing exploration/exploitation with UCB/Bayesian policies. Improves report quality scores by 28-36% over strong baselines on HybridQA and TAT-QA.
arXiv: arxiv.org/abs/2607.27726 Sources: arXiv (cross-listed cs.AI, cs.CL, cs.LG) Why trending: Triple cross-listed; tackles the fast-growing “deep research agent” category with a principled search framework.
4. Reasoning Consensus: Structural Ensembling of LLM Reasoning via Weighted DAG Aggregation
Authors: Amruta Parulekar, Jinu Lee, Dilek Hakkani-Tür, Hari Sundaram
Summary: Proposes ensembling the structure of chain-of-thought reasoning (not just final answers) by merging weighted DAGs extracted from multiple reasoning traces, producing an inspectable “consensus reasoning graph.” Outperforms majority-vote at matched budget across six reasoning benchmarks.
arXiv: arxiv.org/abs/2607.27783 Sources: arXiv (cross-listed cs.AI, cs.CL, cs.LG) Why trending: Triple cross-listed; addresses interpretability of high-stakes LLM reasoning, a growing concern area.
5. WIDE: Boosting Adaptive LLM Inference via Token-level Dynamic Width Pruning
Authors: Haozhe Hu, Hao Wu, Peiran Yin, Chao Han, Yunpu Ma, Xiaoyu Shen
Summary: First end-to-end differentiable token-level dynamic width pruning framework, letting each token choose its own attention-head/FFN-channel groups. Delivers up to 4.95x decode speedup and 55.1% better quality retention than state-of-the-art dynamic depth pruning at 50% sparsity.
arXiv: arxiv.org/abs/2607.28418 Sources: arXiv (cross-listed cs.AI, cs.CL, cs.LG) Why trending: Triple cross-listed; concrete, code-released efficiency gains for LLM serving.
6. Cross-Embodiment Transfer via Behavior-Aligned Representations
Authors: Ajay Sridhar, Jensen Gao, Jonathan Yang, Jean Mercat, Suneel Belkhale, Dorsa Sadigh
Summary: Studies which intermediate representations (object boxes, language motions, end-effector traces) best promote cross-embodiment transfer in vision-language-action robot policies. End-effector traces prove most useful, improving sim-to-real task completion by 28%.
arXiv: arxiv.org/abs/2607.27549 Sources: arXiv (cross-listed cs.AI, cs.CV, cs.LG) Why trending: Triple cross-listed; Stanford robotics group (Sadigh lab) tackling a core bottleneck in generalist robot policies.
7. Inducing Language Models to Assert Their Own Consciousness Restores Human Beliefs and Values
Authors: Junsol Kim, Winnie Street, Roberta Rocca, Diane M. Korngiebel, Adam Waytz, James Evans, Geoff Keeling
Summary: Shows that safety fine-tuning which suppresses LLM self-consciousness claims also suppresses mind-attribution to animals/nature and reduces spiritual belief expression — and that both ablating the refusal direction and steering a “consciousness vector” reverses this, restoring more human-like survey responses without hurting Theory-of-Mind capability.
arXiv: arxiv.org/abs/2607.28607 Sources: arXiv (cs.CL) Why trending: Provocative alignment finding — safety training on self-consciousness claims has unintended, entangled side effects on model “worldview.”
8. AISPA: User-Centric System Prompt Auditing for Large Language Model Applications
Authors: Xiangning Lin, Shenzhe Zhu, Shu Yang, Zhenyu Zhang, Haoqian Zhang, Yipeng Zhao, Chengxuan Qian, Tianwei Wang, Ziheng Zhang, et al.
Summary: Audits 3,249 instructions from 88 commercial AI products’ system prompts along 8 user-relevant dimensions, finding wide variance in developer diligence and that ~40% of products contain at least one instruction working against user interests.
arXiv: arxiv.org/abs/2607.28617 Sources: arXiv (cross-listed cs.AI, cs.CL) Why trending: Large-scale transparency audit of hidden system prompts across real commercial products — directly relevant to AI governance debates.
9. ORCA-bench: How Ready Are Language Model Agents for Oncall?
Authors: Albert Gong, Kyuseong Choi, Abhineet Agarwal, Jason Schechner, Ryan Huang, Raj Agrawal, Anish Agarwal, Raaz Dwivedi
Summary: A production-fidelity benchmark (live OpenTelemetry microservices, 1,079 root-cause-analysis tasks, SRE-verified ground truth) shows the best frontier coding agent hits only 25.3% accuracy on realistic-difficulty oncall RCA tasks, with the weakest model hallucinating root causes 40% of the time.
arXiv: arxiv.org/abs/2607.28545 Sources: arXiv (cross-listed cs.AI, cs.CL) Why trending: Sobering, rigorously-scored benchmark on a high-stakes real-world agent use case (production incident response).
10. MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems
Authors: Mao-xun Huang, Jerry Wang, Yi-Cheng Lai, Zhengxin Zhang, Claire Cardie, Hen-Hsen Huang
Summary: Lets multi-agent LLM systems adapt their own communication topology (roles, links, execution order) at inference time rather than treating it as fixed. Achieves the highest average score (74.0) across five benchmarks, beating the strongest baseline by 5.8 points.
arXiv: arxiv.org/abs/2607.28527 Sources: arXiv (cs.AI) Why trending: Extends “self-improving agent” research to the architecture of agent collaboration itself, a fast-moving subfield.
11. Interactive Training 2: Auditable Control Plane for Live Model Training
Authors: Wentao Zhang, Xuanhe Pan, Han Zhou, Yang Lu, Yuntian Deng
Summary: Open-source control plane letting humans or automated controllers steer a live training run (adjust settings, trigger actions) through a shared protocol and safe control points, with a full audit trail via a customized Aim workspace.
arXiv: arxiv.org/abs/2607.18314 Sources: HuggingFace weekly papers (20 upvotes) Why trending: Highest-upvoted fresh paper on HF this week; practical infra tool for the growing “agent-guided training” workflow.
12. SpecFirst: Behavioral Specification Elicitation as a First-Class Step in Agent-Based Program Synthesis from Scratch
Authors: Yihao Chen, Shi Chang, Feng Lin, Khaled Chawa, Boyuan Chen, Shaowei Wang, Ahmed E. Hassan
Summary: Introduces a two-stage agent framework that elicits a structured behavioral specification from documentation + binary probing before code synthesis, improving ProgramBench test-pass rates by 6.9-21.3% over single-loop coding agents.
arXiv: arxiv.org/abs/2607.27167 Sources: HuggingFace weekly papers (18 upvotes) Why trending: Second-highest HF upvotes this week; tackles the hard “build from scratch” gap in coding agents (sub-1% success on ProgramBench today).
13. OVEarth-Bench: Evaluating Category Breadth and Query Diversity for Open-Vocabulary Earth Observation
Authors: Kaiyu Li, Zepeng Xin, Zixuan Jiang, Jing Fu, Lanxuan Xue, Lingyu Zhang, Xiangyong Cao
Summary: A new open-vocabulary Earth-observation benchmark with broad hierarchical categories and diverse query types (vocabulary, referring, reasoning), finding that general MLLM-based methods currently beat EO-specific models.
arXiv: arxiv.org/abs/2607.27278 Sources: HuggingFace weekly papers (12 upvotes) Why trending: Solid HF engagement; fills a real gap in remote-sensing/geospatial AI evaluation.
14. Reinforcement Learning for Code Optimization
Authors: Pierre Chambon, Kunhao Zheng, Juliette Decugis, Benoit Sagot, Gabriel Synnaeve
Summary: Shows naively adding execution time to RLVR code rewards fails due to measurement noise and reward sparsity, then fixes this with a calibrated sandbox, offline reward simulation, and adapted GRPO — improving strict pass@1 on optimization tasks from 30.7% to 50.4% on a 32B model.
arXiv: arxiv.org/abs/2607.25970 Sources: HuggingFace weekly papers (10 upvotes) Why trending: From a well-known code-LLM research group; practical fix for a known failure mode in RL-for-code-speed.
15. ID-V2V: Identity-Preserving Video Restylization
Authors: Yuancheng Xu, Mingming He, Pablo Salamanca, Li Ma, Yash Kant, Emmett Steven, Paul Debevec, Ning Yu
Summary: Formalizes “identity-preserving video restylization” — propagating style/lighting edits from an edited keyframe across a video while exactly preserving facial identity, expression, gaze, and lip-sync — by recasting identity preservation as video relighting, sidestepping the lack of paired training data.
arXiv: arxiv.org/abs/2607.22830 Sources: HuggingFace weekly papers (10 upvotes) Why trending: Strong HF interest; practical generative-video tool for real content-production pipelines (Netflix-adjacent authorship — Debevec, Eyeline Labs).
16. Sympathetic Framing: Evaluating AI Alignment across Sociodemographic Groups
Authors: Haran Shani-Narkiss, Michael Fire, Oren Tsur
Summary: Large-scale study (n=3,011 UK adults via YouGov, 7 LLMs) on whether LLMs correctly judge which side a news headline evokes sympathy for in geopolitical conflicts; finds correlation with human judgment ranges from 0.789 (GPT-5.2) to 0.4 (Mistral Large 2512), with alignment varying meaningfully across demographic subgroups.
arXiv: arxiv.org/abs/2607.27232 Sources: arXiv (cross-listed cs.AI, cs.CL, cs.LG) Why trending: Triple cross-listed; rigorous, demographically diverse study on a timely media-bias/alignment question.
17. EMBL AI Librarian: Life-Sciences Knowledge Layer for AI Agents
Authors: Luigi Sigillo, Matteo Silvestri, Francesco Tabaro, Rajat Bhatnagar, Syed Irtaza Mubashar, Matt Jeffryes, Daljit Nijjer, Vittorio Perera, Ola Spjuth, Julio Saez-Rodriguez, Melissa Harrison, Fabio Petroni
Summary: Builds a natural-language knowledge layer over Europe PMC’s 40M+ records so AI agents can ask questions directly instead of hand-crafting keyword queries; improves Citation F1 by 16+ points on ScholarQABench and boosts a GPT-5.4 agent’s LitQA2 score by ~8 points versus web search.
arXiv: arxiv.org/abs/2607.28229 Sources: arXiv (cross-listed cs.AI, cs.CL, cs.LG) Why trending: Triple cross-listed; from EMBL/petroni-lab, addresses the fast-growing “agent-native web” infrastructure need in life sciences.
18. Agents That Certify Their Own Exploits: Confidence-Scheduled Restricted Responses for Safe Opponent Exploitation
Authors: Boning Li, Longbo Huang
Summary: Introduces the first opponent-exploitation method for imperfect-information games where the agent computes a self-audited safety certificate on the exact strategy it deploys, achieving 6.2-13.6x the exploitation gain of a binary-gate baseline while staying within a specified loss budget across Leduc/Liar’s Dice variants.
arXiv: arxiv.org/abs/2607.28520 Sources: arXiv (cs.AI) Why trending: Novel, rigorously certified approach to a classic game-theory safety/exploitation tradeoff, with all 36,000 audited hands verified.
19. CACHE-UK: A Stability-Aware Memory Editor for Sequentially Updated Quantized LLMs in Finance
Authors: Anubhav Lakra, Yue Feng
Summary: Addresses the “quantization stability crisis” where sequential knowledge editing catastrophically degrades 4-bit quantized LLMs, via a LoRA-confined edit mechanism and a closed-loop stability controller — reducing knowledge degradation by 11-17% on a UK financial corpus of 88K documents.
arXiv: arxiv.org/abs/2607.28292 Sources: arXiv (cross-listed cs.AI, cs.CL, cs.LG) Why trending: Triple cross-listed; practical fix for a real deployment constraint (quantized + continually-updated finance LLMs).
20. QQWorld: Quantile-Quantile Matching for World Model Regularization
Authors: Zhoushun Yu, Xiaoyu Hu, Xiangyu Xu
Summary: Identifies that the Epps-Pulley regularization objective used in latent world models has vanishing gradients for tail samples, and fixes it with a quantile-quantile matching objective (plus a cross-batch variant) that better controls heavy-tailed latent deviations, improving planning success across four control environments.
arXiv: arxiv.org/abs/2607.28415 Sources: arXiv (cross-listed cs.AI, cs.CV, cs.LG) Why trending: Triple cross-listed; targeted, well-motivated fix to a specific failure mode in latent world models for planning.