Daily AI Papers — August 01, 2026

1. DistillAlign: Coordinating Mode Covering and Mode Seeking in Autoregressive Video Distillation

Authors: Jiaxing Li, Kai Zou, Cindy Zhou, Kaichen Huang, Junyao Gao, Zile Wang, Yang Liu, Bin Liu, Bo An, Yangguang Li

Summary: Revisits the standard multi-stage Distribution Matching Distillation pipeline for autoregressive video distillation, arguing that decoupling the initialization stage from the DMD stage causes them to pursue mismatched target distributions. DistillAlign proposes a coordinated mode-covering/mode-seeking scheme so the initialized student better matches what DMD actually optimizes for, rather than being judged only on generic visual scores like VBench.

arXiv: arxiv.org/abs/2607.26811 Sources: HuggingFace Daily Papers (88 upvotes on 2026-07-30), HuggingFace Trending board Why trending: Highest upvote count among unfeatured papers this cycle; addresses a known blind spot in the widely-used DMD video distillation recipe.

2. MindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program Synthesis

Authors: Yihao Chen, Shi Chang, Khaled Chawa, Feng Lin, Boyuan Chen, Shaowei Wang, Ahmed E. Hassan

Summary: Tackles from-scratch program synthesis — a much harder setting than bug-fixing or feature addition, where even frontier models solve under 1% of ProgramBench tasks. MindForge builds scalable, source-free training environments spanning the whole software-engineering lifecycle so small language models can learn end-to-end synthesis rather than just patching existing code.

arXiv: arxiv.org/abs/2607.27146 Sources: HuggingFace Daily Papers (22 upvotes on 2026-07-30) Why trending: From-scratch code generation remains one of the biggest unsolved gaps for coding agents, and a “teach small models” angle is attractive for cost-conscious deployment.

3. β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

Authors: Jiawei Xu, Minghui Liu, Juzheng Zhang, Tom Goldstein, Furong Huang

Summary: Identifies that vanilla on-policy self-distillation (OPSD) — a promising but brittle technique for improving reasoning LLMs — is just the β=1 case of a broader policy-optimization family, where β weights the KL penalty to a reference policy. Turning β into a tunable hyperparameter rather than an implicit constant makes OPSD training substantially more reliable.

arXiv: arxiv.org/abs/2607.28582 Sources: HuggingFace Daily Papers (17 upvotes on 2026-07-31) Why trending: Practical fix for a known pain point in reasoning-model post-training, backed by well-known RL/ML researchers (Tom Goldstein, Furong Huang).

4. Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Authors: Peilin Feng, Suorong Yang, Soujanya Poria

Summary: Points out that existing agent memory systems record interaction content but not which peer agents are trustworthy under which conditions. Σ-Mem introduces an online memory that tracks historical competence evidence and peer-relationship evidence, letting a central agent judge plausible-but-unverifiable peer responses in multi-agent settings.

arXiv: arxiv.org/abs/2607.27958 Sources: HuggingFace Daily Papers (12 upvotes on 2026-07-31) Why trending: Multi-agent reliability and trust modeling is a fast-growing niche as agent swarms get deployed for real tasks.

5. INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

Authors: Junhan Sun, Hao Zhao, Guofeng Zhang

Summary: Forward latent world models can predict how actions change a scene but need expensive test-time search to recover the action for a desired change. INTACT is an end-to-end JEPA-style model that learns an isomorphic intent-to-action interface directly from action-labeled, reward-free trajectories, removing the need for search at deployment.

arXiv: arxiv.org/abs/2607.26056 Sources: HuggingFace Daily Papers (12 upvotes on 2026-07-31), HuggingFace Trending board Why trending: Search-free world-model control is a hot direction following the broader JEPA/world-model research wave.

6. Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Authors: Alexi Gladstone, Heng Ji, Yilun Du

Summary: Argues that generative modeling has never fully embraced end-to-end training the way discriminative deep learning did, because handling multi-modal distributions has forced generation to be factored into hand-designed stages. The paper proposes “explorative modeling” as a third pretraining axis intended to unlock genuinely end-to-end generative training.

arXiv: arxiv.org/abs/2607.27372 Sources: HuggingFace Daily Papers (11 upvotes on 2026-07-31) Why trending: A foundational, provocative reframing of generative model training from a recognizable trio of authors (Heng Ji, Yilun Du).

7. LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

Authors: Enjun Du, Hange Zhou, Chenxu Du, Siyi Liu, Zirong Chen, Ziyu Zheng, Yongqi Zhang

Summary: Notes that multimodal VQA agents are evaluated almost entirely on final-answer accuracy, which can’t distinguish a grounded correct answer from one reached via language priors or lucky error cancellation. LEDGERMIND treats an agent’s trajectory as a provenance-constrained state machine, normalizing tool outputs into a structured evidence ledger to audit how conclusions were actually reached.

arXiv: arxiv.org/abs/2607.28374 Sources: HuggingFace Daily Papers (10 upvotes on 2026-07-31) Why trending: Addresses growing concern over shallow “final-answer-only” evaluation of multimodal agentic reasoning.

8. Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

Authors: Yash Pandya, Sahil Gupta, Sarthak Harne, Archana Yadav, Kavyansh Chourasia, Hussein Mozannar, Vibhav Vineet, Sara Abdali, Corby Rosset, Yash Lara, Ahmed Awadallah, Ece Kamar, Akshay Nambi

Summary: Argues that as synthetic-environment generation pipelines scale up in quantity, the bottleneck shifts from “how many environments exist” to “how much behavioral depth is inside each one.” Echoverse focuses on building deep, evolving, stateful application environments — including login-gated ones — to train computer-use agents more effectively.

arXiv: arxiv.org/abs/2607.28074 Sources: HuggingFace Daily Papers (9 upvotes on 2026-07-31) Why trending: Microsoft Research authorship (Mozannar, Awadallah, Kamar, Nambi) and direct relevance to the computer-use-agent training bottleneck everyone is racing to solve.

9. Multi-Head Attention Residuals

Authors: Cheng Luo, Zefan Cai, Junjie Hu

Summary: Observes that standard transformers propagate depth information through a single additive residual stream read by one shared query, forcing every feature subspace to read the same layer-history distribution even when subspaces disagree about which layers matter. The paper proposes per-head attention residuals so each subspace can independently learn which depths to attend to.

arXiv: arxiv.org/abs/2607.27230 Sources: HuggingFace Daily Papers (8 upvotes on 2026-07-31) Why trending: A clean architectural critique of the residual stream, a core transformer primitive that rarely gets revisited.

10. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Authors: Sizhe Zhou, Sheldon Yu, Hui Wei, Junda Wu, Siru Ouyang, Yizhu Jiao, Shijia Pan, Julian McAuley, Yu Zhang, Tong Yu, Jiawei Han

Summary: Points out that many deployed LLM agents already use plain filesystems (markdown directory trees) as long-term memory, yet research has largely ignored this default in favor of bespoke memory architectures. The paper stress-tests whether agents can actually keep such a store organized as memories accumulate, conflict, and go stale over time.

arXiv: arxiv.org/abs/2607.26637 Sources: HuggingFace Daily Papers (6 upvotes on 2026-07-31) Why trending: Filesystem-as-memory is the de facto approach in many real agent products (e.g. Claude/Codex-style agents), making this an unusually practical memory paper.

11. Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

Authors: Tianyu Wang, Yuxuan Zhou, Wenbin Wang, Heng Li, Zikai Xiao, Junyuan Shang

Summary: Examines lossy verification schemes in speculative decoding, which relax strict distributional matching between draft and target models to gain more speedup. The paper shows this relaxation silently rewrites the decoding distribution and can cause unstable or severely degraded generation quality, mapping out the underlying mechanisms and failure modes.

arXiv: arxiv.org/abs/2607.26627 Sources: HuggingFace Daily Papers (4 upvotes on 2026-07-31) Why trending: Speculative decoding is now standard in production LLM serving, and quality regressions from “faster” lossy variants are an active operational concern.

12. Harness-G: A Graph-Structured Harness for Search Agents

Authors: Yanning Hou, Haoyuan Chen, Sihang Zhou, Xiaoshu Chen, Xirui Liu, Duanyang Yuan, Lingyuan Meng, Quan Liu, Jian Huang

Summary: Identifies “retrieval aliasing” in RL-trained search agents (e.g. Search-R1 style): rollouts for the same question generate distinct free-form query strings that nonetheless retrieve near-identical documents, muddying credit assignment. Harness-G restructures the policy-environment interface as a graph to give retrieval a properly structured formulation.

arXiv: arxiv.org/abs/2607.27652 Sources: HuggingFace Daily Papers (3 upvotes on 2026-07-31) Why trending: Diagnoses a subtle but consequential bug in the popular Search-R1 RL-for-retrieval training paradigm.

13. OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models

Authors: Jinsen Su, Yongdong Luo, Yuexiao Ma, Yibo Hu, Meiguang Jin, Xiaowu Zheng

Summary: Shows that existing token-compression methods for omnimodal LLMs assume one modality (e.g. video) can dictate what’s relevant in another (e.g. audio), but audio/video relevance for the same query often peaks at different moments. OmniScope is a training-free compression framework that lets the query independently steer retention per modality.

arXiv: arxiv.org/abs/2607.23193 Sources: HuggingFace Daily Papers (3 upvotes on 2026-07-31) Why trending: Training-free efficiency gains for omnimodal models are highly practical as audio-video LLMs proliferate.

14. Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions

Authors: Pengyu Zhu, Lijun Li, Longju Yang, Sen Su, Jing Shao

Summary: Tests whether “Deep Research” agents — which plan, retrieve, and synthesize reports over long horizons — can resist plausible-looking but factually false information planted in their sources. The authors build MisKnow-Agent, a controlled evaluation framework with manually audited false conclusions and controlled authority/source-style cues, and find current deep-research agents are notably susceptible.

arXiv: arxiv.org/abs/2607.20891 Sources: HuggingFace Daily Papers (3 upvotes on 2026-07-31) Why trending: Direct, timely critique of the “Deep Research” agent category that OpenAI, Google, and others have heavily marketed.

15. AI Tour Meeting: Group Travel Planning by LLM Agents

Authors: Daisuke Kikuta

Summary: Proposes a multi-agent group travel planning framework where LLM agents are instantiated with distinct personas representing different travelers and negotiate an itinerary through natural-language discussion. The framework provides configurable interfaces for persona setup, discussion workflow, and monitoring.

arXiv: arxiv.org/abs/2607.18806 Sources: HuggingFace Daily Papers (3 upvotes on 2026-07-31) Why trending: A concrete, relatable multi-agent negotiation application that showcases persona-driven agent collaboration.

16. ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Authors: Yao Xiao, Reuben Tan, Zhen Zhu, Yuqun Wu, Jianfeng Gao, Derek Hoiem

Summary: Tackles the problem that vision-language model performance degrades as visual distractors increase, while processing all visual tokens at once is infeasible under GPU memory limits. ReToken adds a single learnable embedding trained as an explicit retrieval target that sparsely selects query-relevant tokens from a pre-filled visual KV cache, trained on just a small image-QA dataset.

arXiv: arxiv.org/abs/2607.28627 Sources: HuggingFace Daily Papers (2 upvotes on 2026-07-31) Why trending: A remarkably lightweight fix (one token) for long visual-context retrieval, with Jianfeng Gao and Derek Hoiem among the authors.

17. Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations

Authors: Pere Martra, Eugenio Martínez Cámara, Alfonso Ureña López

Summary: Introduces Fairness Pruning, a lightweight structural intervention for locating and eventually mitigating demographic bias in LLMs by identifying neurons in GLU-MLP layers that react differentially to demographic attributes. Uses minimally contrastive prompt pairs and inference-time activation capture for causal bias localization.

arXiv: arxiv.org/abs/2607.28319 Sources: HuggingFace Daily Papers (2 upvotes on 2026-07-31) Why trending: Interpretability-driven approach to bias mitigation is an area of steady interest for responsible-AI research.

18. Beyond Geometric Complementarity: Coherent Overlap in Sparse Mixture-of-Experts Routing

Authors: Huiyuan Tian, Bonan Xu, Shijian Li

Summary: Challenges the common geometric explanation for why sparse MoE routing helps — that co-selected experts contribute distinct representation directions — by disentangling route coherence, candidate quality, and candidate-context interaction using a new Expert Subspace Separation Index and controlled interventions.

arXiv: arxiv.org/abs/2607.28308 Sources: HuggingFace Daily Papers (2 upvotes on 2026-07-31) Why trending: Rigorous mechanistic analysis of MoE routing behavior, relevant given how central MoE architectures have become to frontier models.

19. AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition

Authors: Yuqi Li, Yi-Cheng Lin, Xianglong Wang, Kuo Yang, Xiaoqin Feng, Yixuan Wang, Huiran Duan, Yingli Tian

Summary: Targets on-device speech emotion recognition, where large self-supervised models perform well but are too costly for edge deployment. AMRD combines adaptive multi-teacher distillation (weighting unreliable teachers via a one-class SVM) with relational, inter-sample distillation rather than plain logit matching.

arXiv: arxiv.org/abs/2607.25289 Sources: HuggingFace Daily Papers (0 upvotes on 2026-07-31, but curated into HF’s daily list) Why trending: Practical edge-deployment recipe for a widely-needed but underserved on-device task.

20. Pedestrian Archetypes Extension — More Pedestrian Models for Autonomous Vehicle Safety Testing

Authors: Taorui Huang, Namita Gaidhani, Ritvik Bansal, S M Jubaer, Regina Lim, Rhett Zhao, Gavin Rafael Selin, Sunnie Deng Gao, Hasnain N Syed

Summary: Extends the authors’ prior “Pedestrian Archetypes” work — which defined 12 behavior-based pedestrian types (Wanderer, Drunk, Distracted, Jaywalker, Elderly, Kid, etc.) — with additional archetype models for more realistic autonomous-vehicle safety testing, moving beyond single-behavior-label pedestrian simulation.

arXiv: arxiv.org/abs/2607.16922 Sources: HuggingFace Daily Papers (0 upvotes on 2026-07-31, but curated into HF’s daily list) Why trending: Practical, safety-critical simulation work for autonomous vehicle testing pipelines.