Daily AI Papers — July 03, 2026

1. Program-as-Weights: A Programming Paradigm for Fuzzy Functions

Authors: Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber, Yuntian Deng (University of Waterloo)

Many everyday programming tasks resist clean rule-based implementation (e.g., alerting on important log lines, repairing malformed JSON) and are increasingly outsourced to LLM APIs at the cost of locality and reproducibility. This paper introduces Program-as-Weights (PAW), where a 4B compiler trained on the 10M-example FuzzyBench dataset emits parameter-efficient adapters for a frozen lightweight interpreter, enabling offline execution at 30 tokens/s on a MacBook M3.

arXiv: arxiv.org/abs/2607.02512 Sources: HuggingFace Daily Papers (#1, 46 upvotes), GitHub (96 stars) Why trending: Proposes a new programming paradigm that compiles natural-language function specs into compact local neural artifacts, enabling a 0.6B model to match Qwen3-32B at 1/50th the memory — huge practical implications for offline/on-device AI.

2. EvoPolicyGym: Evaluating Autonomous Policy Evolution in Interactive Environments

Authors: Zhilin Wang, Han Song, Runzhe Zhan, Jusen Du, Jiacheng Chen, Tianle Li et al.

This paper introduces Autonomous Policy Evolution, a controlled evaluation where a model agent repeatedly edits an executable policy system under a fixed interaction budget. EvoPolicyGym provides trajectory-level diagnostics that reveal how agents allocate budget and convert feedback into parametric tuning, showing that strong autonomous policy evolution depends on discovering task-appropriate mechanisms.

arXiv: arxiv.org/abs/2607.02440 Sources: HuggingFace Daily Papers (#2, 39 upvotes) Why trending: First benchmark to evaluate how well agents iteratively improve executable RL policies under fixed budgets — GPT-5.5 leads across all 16 environments, sparking community debate about agent evaluation methodology.

3. AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents

Authors: Xiangchen Cheng, Yunwei Jiang, Jianwen Sun, Zizhen Li, Chuanhao Li, Kaipeng Zhang et al.

Traditional long-horizon agent evaluations append all past observations to prompts, creating an entangled mixture where the effect of any single memory component is hard to isolate. AgenticSTS instruments a bounded-context alternative with structured memory slots to enable systematic comparison of memory strategies at scale.

arXiv: arxiv.org/abs/2607.02255 Sources: HuggingFace Daily Papers (#3, 38 upvotes) Why trending: Addresses a critical gap in long-context agent evaluation by isolating the effect of individual memory components under a bounded-context contract, enabling controlled studies of memory architecture choices.

4. Morphing into Hybrid Attention Models

Authors: Disen Lan, Jianbin Zheng, Yuxi Ren, Xin Xia, Xuanda Wang et al.

Hybrid attention models improve long-context efficiency by retaining only a subset of full-attention layers while replacing others with linear attention. This paper proposes a learned layer-selection method that outperforms existing heuristic strategies (fixed placement patterns), enabling better Transformer-to-hybrid conversion for production deployments.

arXiv: arxiv.org/abs/2606.30562 Sources: HuggingFace Daily Papers (#4, 30 upvotes) Why trending: Provides a principled, data-driven approach for converting full-attention Transformers into efficient hybrid models — directly relevant to anyone deploying long-context LLMs at scale.

5. AgenticDataBench: A Comprehensive Benchmark for Data Agents

Authors: Zhaoyan Sun, Shan Zhong, Daizhou Wen, Jiaxing Han, Guoliang Li et al.

LLM-based data agents have emerged as a promising solution for automating data science tasks over heterogeneous raw data. AgenticDataBench provides a comprehensive evaluation suite spanning data ingestion, transformation, analysis, and visualization tasks, establishing rigorous baselines for comparing state-of-the-art data agents.

arXiv: arxiv.org/abs/2607.01647 Sources: HuggingFace Daily Papers (#5, 20 upvotes) Why trending: First comprehensive benchmark specifically targeting LLM-based data agents across the full data science pipeline, filling a major gap as data agents go mainstream.

6. Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling

Authors: Xingyu Zheng, Xianglong Liu, Yifu Ding, Weilun Feng, Junqing Lin et al.

This paper presents a training-free acceleration strategy for flow matching diffusion models using multi-resolution staged sampling, performing coarse generation at low resolution before progressive upsampling. Unlike timestep distillation methods, it requires no retraining and attains hardware-agnostic speedups across standard diffusion pipelines.

arXiv: arxiv.org/abs/2607.01642 Sources: HuggingFace Daily Papers (#6, 19 upvotes) Why trending: Achieves >5× speedup on text-to-image diffusion without any training or custom kernels, with practical multi-resolution staged sampling that works out-of-the-box on existing models.

7. WorldDirector: Building Controllable World Simulators with Persistent Dynamic Memory

Authors: Hanlin Wang, Hao Ouyang, Qiuyu Wang, Wen Wang, Qingyan Bai et al.

WorldDirector is a highly controllable video world model framework that explicitly separates semantic motion orchestration from visual generation, unlike existing models that entangle physical dynamics with pixel rendering. It maintains persistent dynamic object memory, allowing consistent object behavior across long video sequences and arbitrary viewpoint changes.

arXiv: arxiv.org/abs/2607.02517 Sources: HuggingFace Daily Papers (#7, 16 upvotes) Why trending: Decouples semantic motion orchestration from pixel rendering in video world models, enabling persistent dynamic object memory and unrestricted viewpoint exploration — a step toward true controllable world simulators.

8. Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning

Authors: Junha Jung, Minbyul Jeong, Suhyeon Lim, Sungwook Jung, Jaehoon Yun et al.

Existing post-training pipelines for medical multimodal LLMs rely on outcome-centric rewards, suffering from sparse credit assignment that makes it difficult to optimize multi-step clinical reasoning. This paper introduces step-aware RL that identifies and penalizes failure cascades — where early reasoning errors propagate and compound — dramatically improving diagnostic reasoning reliability.

arXiv: arxiv.org/abs/2606.31825 Sources: HuggingFace Daily Papers (#8, 13 upvotes) Why trending: Directly targets the medical AI safety concern of cascading reasoning errors in clinical image analysis, with step-level RL credit assignment that improves reliability in high-stakes settings.

9. SkillCoach: Self-Evolving Rubrics for Evaluating and Enhancing Agentic Skill-Use

Authors: Jiayin Zhu, Kelong Mao, Yudong Guo, Dengbo He, Sulong Xu et al.

As LLM agents increasingly rely on large libraries of reusable skills (SOPs, tool workflows, domain rules), overlapping skills make reliable skill-use difficult and final-answer verification is too coarse for training. SkillCoach generates and iteratively refines step-level evaluation rubrics, enabling agents to self-improve on skill selection and execution without human annotation.

arXiv: arxiv.org/abs/2607.01874 Sources: HuggingFace Daily Papers (#9, 12 upvotes) Why trending: Addresses the emerging problem of agents failing to correctly select and execute skills in large skill repositories, with self-evolving rubrics that improve both evaluation and training signals.

10. Optimizing Visual Generative Models via Distribution-wise Rewards

Authors: Ruihang Li, Mengde Xu, Shuyang Gu, Leigang Qu, Fuli Feng et al.

Conventional RL strategies for visual generation employ sample-wise reward functions, frequently resulting in reward hacking that degrades image diversity and introduces visual anomalies. This paper presents a framework that finetunes generative models using distribution-wise rewards, ensuring better alignment with real-world image statistics while preventing mode collapse.

arXiv: arxiv.org/abs/2607.02291 Sources: HuggingFace Daily Papers (#10, 12 upvotes) Why trending: Directly addresses reward hacking in RLHF for image generation — a widely known failure mode — with a distribution-level reward framework that preserves diversity while improving alignment.

11. Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads

Authors: Aryo Pradipta Gema, Beatrice Alex, Pasquale Minervini

In long-context LLMs, models often synthesize answers from the meaning of relevant spans rather than literally copy-pasting them, but existing retrieval head detectors miss these by only rewarding literal token matches. Logit-Contribution Scoring identifies non-literal retrieval heads by measuring how much each head’s attended tokens contribute to the final output distribution, revealing a new class of semantically-active heads.

arXiv: arxiv.org/abs/2607.01002 Sources: HuggingFace Daily Papers (#11, 10 upvotes) Why trending: Advances mechanistic interpretability of long-context LLMs by discovering attention heads responsible for semantic synthesis rather than literal copying — important for understanding and improving RAG systems.

12. When Search Agents Should Ask: DiscoBench for Clarification-Aware Deep Search

Authors: Yiling Tao, Shihan Deng, Meiling Tao, Pengzhi Wei, Zhichao Hu et al.

Existing search agent benchmarks assume queries are complete and explicit, ignoring real-world scenarios where requests are vague, underspecified, or factually incorrect. DiscoBench evaluates clarification-aware deep search, measuring agents’ ability to detect query ambiguity and proactively request clarification before investing in multi-step retrieval.

arXiv: arxiv.org/abs/2606.27669 Sources: HuggingFace Daily Papers (#16, 7 upvotes) Why trending: Targets the critical but underexplored problem of agents knowing when to ask for clarification vs. plowing ahead with underspecified or incorrect queries — essential for real-world search agent deployment.

13. PACE: A Proxy for Agentic Capability Evaluation

Authors: Yueqi Song, Lintang Sutawika, Jiarui Liu, Lindia Tjuatja, Jiayi Geng et al.

Evaluating LLM agents on SWE-Bench and GAIA can cost thousands of dollars and take days per run. PACE investigates whether performance on fast, cheap non-agentic capability benchmarks (reasoning, code generation) reliably predicts agentic task performance, providing a practical shortcut for rapid iteration during model development.

arXiv: arxiv.org/abs/2607.02032 Sources: HuggingFace Daily Papers (#17, 4 upvotes) Why trending: Proposes cheap proxy benchmarks that predict expensive agentic benchmark performance (SWE-Bench, GAIA) — addresses the real economic pain of $1000+ evaluation runs.

14. Learning to Move Before Learning to Do: Task-Agnostic Pretraining for VLAs

Authors: Junhao Shi, Siyin Wang, Xiaopeng Yu, Li Ji, Jingjing Gong et al.

Vision-Language-Action (VLA) models are bottlenecked by the scarcity of expert demonstrations, conflating two distinct learning objectives: acquiring physical competence (how to move) and semantic alignment (what to do). This paper shows that task-agnostic motor pretraining on cheap, unlabeled motion data can be learned independently, dramatically reducing the expert demonstrations needed for downstream task learning.

arXiv: arxiv.org/abs/2607.02466 Sources: HuggingFace Daily Papers (#19, 4 upvotes) Why trending: Proposes a principled separation of motor pretraining from semantic alignment in robotics — could drastically reduce the need for costly expert demonstrations in robot learning.

15. Denser ≠ Better: Limits of On-Policy Self-Distillation for Continual Post-Training

Authors: Meng Wang, Haohan Zhao, Wenzhuo Liu, Lu Yang, Geng Liu et al.

Continual post-training enables foundation models to acquire new knowledge while preserving existing capabilities, and on-policy self-distillation has emerged as an attractive approach. This work revisits this via self-distillation policy optimization (SDPO) experiments, revealing that denser supervision does not reliably prevent catastrophic forgetting and can hurt generalization under distribution shift.

arXiv: arxiv.org/abs/2607.01763 Sources: HuggingFace Daily Papers (#20, 4 upvotes) Why trending: Challenges the popular optimism around on-policy self-distillation for continual learning, showing concrete failure modes — timely pushback against a growing trend in LLM post-training.

16. Representation Distribution Matching for One-Step Visual Generation

Authors: Lan Feng, Wuyang Li, Eloi Zablocki, Matthieu Cord, Alexandre Alahi

This paper elucidates the design space of Representation Distribution Matching (RDM), training one-step image generators by matching generated and reference feature distributions under frozen pretrained encoders. Controlled studies along two design axes — how distributions are compared and in which representation space — yield actionable insights for building competitive one-step generators.

arXiv: arxiv.org/abs/2607.02375 Sources: HuggingFace Daily Papers (#18, 4 upvotes) Why trending: Clarifies the design space of one-step image generation via distribution matching under frozen encoders — practically relevant for anyone training fast single-step generators.

17. AGVBench: A Reliability-Oriented Benchmark of Data Augmentation for Vein Recognition

Authors: Haiyang Li, Yuming Fu, Qun Song, Hongchao Liao, Jing Chen et al.

Vein recognition is a secure biometric technology constrained by limited annotated data and imaging variations, and data augmentation strategies designed for natural images may disrupt the fine-grained topology and textures essential for identity discrimination. AGVBench evaluates 30 representative augmentation strategies on five public palm- and finger-vein datasets, providing the community with reliability-oriented baselines.

arXiv: arxiv.org/abs/2607.02271 Sources: HuggingFace Daily Papers (#12, 9 upvotes) Why trending: First systematic benchmark of 30 augmentation strategies for vein biometrics, revealing that standard natural-image augmentations can degrade fine-grained topology essential for identity discrimination.

18. InstanceControl: Controllable Complex Image Generation without Instance Labeling

Authors: Xiaoyu Liu, Huan Wang, Fan Li, Zhixin Wang, Jiaqi Xu et al.

ControlNet-style methods struggle with complex multi-instance scenes, frequently causing attribute confusion among instances, and existing solutions require labor-intensive manual instance labeling. InstanceControl introduces a label-free approach that automatically disentangles instance attributes, enabling reliable controllable generation of scenes with multiple interacting objects.

arXiv: arxiv.org/abs/2606.31924 Sources: HuggingFace Daily Papers (#14, 8 upvotes) Why trending: Eliminates the need for manual instance labeling in multi-object controllable generation, enabling ControlNet-style guidance for complex scenes without annotation overhead.

19. From SRA to Self-Flow: Data Augmentation or Self-Supervision?

Authors: Dengyang Jiang, Mengmeng Wang, Harry Yang, Jingdong Wang

Recent self-alignment methods for diffusion transformers (SRA and Self-Flow) remove the dependency on external pretrained encoders by constructing alignment within the diffusion model itself, but the mechanism behind their improvement is poorly understood. This paper systematically investigates whether the gain comes from data augmentation or the self-supervised learning formulation, providing clarity for practitioners.

arXiv: arxiv.org/abs/2607.02508 Sources: HuggingFace Daily Papers (#13, 9 upvotes) Why trending: Clarifies the mechanism behind recent self-alignment diffusion training improvements (SRA vs. Self-Flow), demystifying a widely adopted but poorly understood technique.

20. AnyGroundBench: A Specialized-Domain Benchmark for Video Grounding in Vision-Language Models

Authors: Rintaro Otsubo, Ryo Fujii, Reina Ishikawa, Taiki Kanaya, Kanta Sawafuji et al.

Current VLM evaluation protocols for spatio-temporal video grounding are largely confined to zero-shot assessments on general daily-life benchmarks, creating a disconnect from real-world applications in specialized fields. AnyGroundBench evaluates models on rare visual concepts and complex spatio-temporal dynamics across specialized domains, revealing significant performance drops that general benchmarks miss.

arXiv: arxiv.org/abs/2607.02269 Sources: HuggingFace Daily Papers (#15, 7 upvotes) Why trending: Exposes a critical gap between VLM video grounding performance on general benchmarks vs. specialized domains, with timely implications for real-world deployment of vision-language systems.