Daily AI Papers — July 28, 2026
Published:
1. Kimi K3: Open Frontier Intelligence
Authors: Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, M. C., et al. (402 authors)
Summary: Kimi K3 is a 2.8T-parameter Mixture-of-Experts model with 104B activated parameters, native vision capability, and a 1M-token context window, built on new “Kimi Delta Attention” and “Stable LatentMoE” (16-of-896 routed experts) architectures. The combination of architectural advances and refined data/training recipes yields roughly a 2.5x improvement in compute efficiency over prior generations.
arXiv: arxiv.org/abs/2607.24653 Sources: HuggingFace Daily Papers (237 upvotes — highest of the day) Why trending: A new open-weight frontier-scale MoE model from Kimi/Moonshot directly competing with closed frontier labs; scale (2.8T params, 1M context) and open release make it the clear top story.
2. JarvisHub: An Open Harness for Canvas-Native Multimodal Creative Agents
Authors: Yunlong Lin, Zixu Lin, Zhaohu Xing, Biqiang Li, Chenxin Li, et al. (26 authors)
Summary: JarvisHub is an open harness for long-horizon, canvas-native creative agents that must manage references, drafts, edits, tool actions, and human feedback across evolving multimodal projects, not just single prompt-output generations. It provides infrastructure for representing this evolving creative state so agents can iterate the way human creators do.
arXiv: arxiv.org/abs/2607.23588 Sources: HuggingFace Daily Papers (100 upvotes) Why trending: Strong community interest in agent harnesses for creative/multimodal generation workflows beyond one-shot image/video synthesis.
3. From Proprietary to Open-Source: Bridging the Distribution Gap via Multi-Agent Protocol Distillation in Agentic Search
Authors: Junlin Liu, Jiangwang Chen, Zixin Song, Shuaiyu Zhou, Chunji Lv, et al. (10 authors)
Summary: The paper tackles the sparse-reward problem in RL-trained agentic search by distilling reasoning/retrieval behavior from proprietary teacher models into open-source students, avoiding the pitfalls of naive logit-matching across differing model families. It proposes a multi-agent protocol-level distillation scheme to densify supervision while preserving the student’s own distribution.
arXiv: arxiv.org/abs/2607.24280 Sources: HuggingFace Daily Papers (63 upvotes) Why trending: Addresses the practical, widely-discussed problem of closing the open-vs-proprietary agent capability gap for search/RAG-style agents.
4. Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation
Authors: Bingnan Li, Haozhe Wang, Haozhong Xiong, Fangtai Wu, Jinpeng Yu, et al. (8 authors)
Summary: The authors show that standard on-policy diffusion distillation (OPD), which naively matches teacher/student CFG-composed velocities, is under-identified at the positive/negative branch level. They propose a corrected formulation that properly disentangles guidance branches during distillation.
arXiv: arxiv.org/abs/2607.24731 Sources: HuggingFace Daily Papers (60 upvotes) Why trending: Fixes a subtle but consequential flaw in a technique widely used to speed up diffusion model inference, relevant to nearly every fast image/video generator.
5. Progress Reward Modeling for Robotic Learning: A Comprehensive Survey
Authors: Jianshu Zhang, Keliang Wu, Haoran Lu, Anbang Liu, Ce Zhang, et al. (11 authors)
Summary: This survey unifies the fragmented literature on “progress rewards” — dense feedback signals that tell a robot whether it is making, stalling, or undoing progress toward a task, unlike terminal-only success signals. It proposes a shared framework for comparing existing methods across observation types and reward formulations.
arXiv: arxiv.org/abs/2607.21655 Sources: HuggingFace Daily Papers (55 upvotes) Why trending: Robotics/embodied-AI reward design is a hot subfield, and a comprehensive survey filling a framework gap draws high community reference value.
6. StateAct: Program State, before Pixels, for Long-Horizon Computer-Use Agents
Authors: Yan Yang, Xiangru Jian, Ziyang Luo, Zirui Zhao, Yutong Dai, et al. (10 authors)
Summary: StateAct argues computer-use agents should reason over underlying program state (files, DOM, backend data) rather than lossy screenshot pixels, since different states can render identically. It introduces a code-first, multi-agent harness where the main agent inspects and manipulates state directly instead of purely visual grounding.
arXiv: arxiv.org/abs/2607.22798 Sources: HuggingFace Daily Papers (51 upvotes) Why trending: Computer-use agents are a major current focus area, and this challenges the pixel-first paradigm most existing agent frameworks follow.
7. Data Pyramid for Embodied Manipulation
Authors: Yifan Ye, Yankai Fu, Yaoxu Lv, Bohan Hou, Jun Cen, et al. (29 authors)
Summary: The paper organizes the embodied-data ecosystem into a five-level “pyramid” (real-robot, UMI-style, ego/exocentric, simulation, and general vision-language data) to clarify how these heterogeneous sources should be combined. It analyzes the tradeoffs of each layer for training generalizable manipulation policies.
arXiv: arxiv.org/abs/2607.24744 Sources: HuggingFace Daily Papers (29 upvotes) Why trending: Data strategy for embodied AI/robotics is a persistent bottleneck topic, and a structured taxonomy is broadly useful to practitioners.
8. Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification
Authors: Haopeng Li, Yitong Li, Junsong Chen, Tian Ye, Haozhe Liu, et al. (11 authors)
Summary: Sol-Attn is a training-free dynamic sparse attention method for video diffusion transformers that addresses the rigid, costly block-routing of prior sparsification approaches. It adaptively selects key-value blocks on the fly to cut inference cost while preserving accuracy on long video token sequences.
arXiv: arxiv.org/abs/2607.24027 Sources: HuggingFace Daily Papers (23 upvotes) Why trending: Video generation inference cost is a major practical pain point, and training-free speedups are immediately adoptable.
9. OmniVAE: An Audio-Video VAE with Cross-Modal Alignment for Joint Generation
Authors: Jun Zhan, Chen Yang, Yitian Gong, Donghua Yu, Kuangwei Chen, et al. (26 authors)
Summary: OmniVAE addresses the misalignment between separately-trained audio and video VAE latent spaces that hampers joint audio-video generation. It introduces cross-modal alignment so a downstream generative model can learn synchronized audio-video correspondence more directly from the shared latent space.
arXiv: arxiv.org/abs/2607.23855 Sources: HuggingFace Daily Papers (20 upvotes) Why trending: Joint audio-video generation with true sync is one of the next frontiers beyond silent video models, drawing strong interest.
10. The Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic Distillation
Authors: Tianyi Men, Zhuoran Jin, Kang Liu, Jun Zhao (4 authors)
Summary: The authors build a controlled multi-turn environment to systematically study how long-horizon planning ability is acquired across pre-training and post-training stages, since standard internet-scale training data is too opaque to isolate this. They analyze single- vs. multi-teacher on-policy agentic distillation as levers for improving planning.
arXiv: arxiv.org/abs/2607.24720 Sources: HuggingFace Daily Papers (19 upvotes) Why trending: Mechanistic, controlled study of agent planning capability acquisition — a rigorous complement to the many black-box agent benchmark papers.
11. Oxygen-TryOn: Fashion-Native Foundation Model for Any-item Virtual Try-On
Authors: Yong Liu, Xiaolong Fu, Zihang Xu, Wen Xue, Xueheng Li, et al. (13 authors)
Summary: Oxygen-TryOn is a foundation model built specifically for virtual try-on (rather than repurposing a general image editor), using a dedicated data engine and try-on-specific training. It can synthesize photorealistic images of a subject wearing arbitrary reference items across fashion categories, going beyond prior systems’ narrower item-type support.
arXiv: arxiv.org/abs/2607.21694 Sources: HuggingFace Daily Papers (19 upvotes) Why trending: Virtual try-on has strong commercial applications in e-commerce, and a purpose-built foundation model (vs. generic editors) is a notable step up.
12. Chamaileon: Cross-Context Binder Design with Contextualized Modeling and Mixed Sampling
Authors: Hengyuan Cao, Shizhuo Cheng, Mingxuan Liu, Weicheng Huang, Yunhong Lu, et al. (8 authors)
Summary: Chamaileon extends protein binder design beyond the common single-target, single-state assumption to model multi-target and multi-state interactions needed for advanced function-oriented design. It combines contextualized sequence-structure modeling with a mixed sampling strategy for generation.
arXiv: arxiv.org/abs/2607.23518 Sources: HuggingFace Daily Papers (6 upvotes) Why trending: AI-for-biology / protein design remains a high-interest cross-over field between ML and structural biology research communities.
13. dRAE: Representation Autoencoder with Hyper-Spherical Codes
Authors: Tianren Ma, Lin Long, Chuyan Chen, Mu Zhang, Junbo Zhao, et al. (7 authors)
Summary: dRAE tackles codebook collapse in visual tokenization by identifying a metric mismatch — standard Euclidean codebook objectives don’t fit the anisotropic geometry of representation space — and instead uses hyper-spherical codes aligned to that geometry. This allows discretizing high-dimensional visual representations to bridge vision and language models without losing semantic coherence.
arXiv: arxiv.org/abs/2607.22148 Sources: HuggingFace Daily Papers (6 upvotes) Why trending: Visual tokenization quality is foundational to multimodal LLMs, and codebook collapse is a well-known unsolved pain point.
14. DecoupleMix: Decoupled Ratio Search and Convex Allocation for Scalable VLM Data Recipes
Authors: Jiahao Xie, Zhongbin Guo, Qianle Wang, Ruiqi Lu, Dongling Xiao, et al. (7 authors)
Summary: DecoupleMix turns VLM pretraining data-mixture design, currently mostly heuristic, into a systematic mixture-optimization problem. It decouples ratio search from convex budget allocation to produce a reproducible, attributable method for constructing pretraining data recipes at scale.
arXiv: arxiv.org/abs/2607.24516 Sources: HuggingFace Daily Papers (5 upvotes) Why trending: Data-mixture optimization is a recurring bottleneck for VLM training that most labs treat as a trade secret; a principled, reproducible method fills a real gap.
15. ClinFusion: A Vision-Centric Multimodal LLM System for Holistic Medical Understanding
Authors: Hangjie Yuan, Yichen Qian, Zhiwei Tang, Xianzhe Xu, Lirong Wu, et al. (24 authors)
Summary: ClinFusion is a vision-centric multimodal LLM designed to absorb knowledge from heterogeneous 2D and 3D medical imaging, aiming at holistic clinical understanding rather than narrow single-modality tasks. It also proposes evaluation protocols aligned with actual radiologist practice for fine-grained, factuality-driven assessment.
arXiv: arxiv.org/abs/2607.24743 Sources: HuggingFace Daily Papers (4 upvotes) Why trending: Medical multimodal AI with clinician-aligned evaluation addresses a recognized weakness (misaligned benchmarks) in healthcare AI research.
16. Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models
Authors: Junlin Fang, Do Nguyen-Thanh, Xiaogang Xu, Zhen Fang, Sean Du (5 authors)
Summary: The paper identifies that long reasoning traces from large reasoning models contain noisy irrelevant and repetitive steps that obscure signals useful for hallucination detection. It proposes a “Reasoning Denoiser” that filters these noisy steps to substantially improve hallucination-detection performance.
arXiv: arxiv.org/abs/2607.22098 Sources: HuggingFace Daily Papers (4 upvotes) Why trending: Hallucination detection in reasoning models (o1/R1-style) is a top safety/reliability concern as reasoning-trace-based models proliferate.
17. Codifying the Judge: Scalable Evaluation via Program Distillation
Authors: Tzu-Heng Huang, Shengqi Qiu, Frederic Sala (3 authors)
Summary: Instead of prompting an LLM-as-judge at evaluation time (costly, slow, opaque), this work distills the judge’s decision logic into a committee of inspectable programs that score candidates directly. This yields transparent, editable, and cheap evaluators as a scalable alternative to LLM-as-judge.
arXiv: arxiv.org/abs/2607.22561 Sources: HuggingFace Daily Papers (4 upvotes) Why trending: LLM-as-judge cost/opacity is a widely-felt pain point in eval pipelines; a program-distillation alternative is immediately practical.
18. FilmBench: A Film-Grade Benchmark for Cinematic Video Generation
Authors: Shengyi Wang, Niantong Li, Guangzheng Hu, Hong Qi, Fei Ding, et al. (30 authors)
Summary: FilmBench critiques existing video-generation benchmarks for relying on web/LLM-templated prompts and generic multimodal scorers rather than professional cinematic-language criteria. It introduces evaluation taxonomies grounded in how films are actually made and judged by industry professionals.
arXiv: arxiv.org/abs/2607.24241 Sources: HuggingFace Daily Papers (3 upvotes) Why trending: As video generation quality approaches professional footage, benchmark rigor grounded in filmmaking craft is a timely and needed contribution.
19. A Frozen 12B Beats Frontier Models on Verified Work: 100% Accuracy, 0 Tokens, Bit-Exact, Forever
Authors: Sietse Schelpe (1 author)
Summary: Instead of retraining models to improve them, this work keeps a 12B model frozen and grows a persistent memory of independently-verified solutions beside it — once a problem family is solved and verified without consulting the answer key, future instances are answered at zero generation tokens, deterministically. Across 180 fresh test instances, this frozen-model-plus-memory approach reportedly matches or beats frontier model accuracy on verified work.
arXiv: arxiv.org/abs/2607.23806 Sources: HuggingFace Daily Papers (3 upvotes) Why trending: A provocative, contrarian claim (no retraining needed, zero-token deterministic answers) against the industry’s retrain-everything norm — the kind of title that draws outsized discussion relative to its author count.
20. GNM Head: A Generative aNthropometric Model of the human head
Authors: Stylianos Ploumpis, Jan Bednarik, Gaspard Zoss, Ruslan Guseinov, Luca Prasso, et al. (29 authors)
Summary: GNM Head is a new parametric/generative model of the human head that extends anatomical scope beyond outer geometry alone to include intra-oral and ocular structures, addressing gaps in existing publicly available head models. It is designed to serve as a conditioning signal for generative vision models needing tight spatial control over head/face imagery.
arXiv: arxiv.org/abs/2607.23687 Sources: HuggingFace Daily Papers (3 upvotes) Why trending: Parametric head models underpin a wide swath of avatar/animation/generative-portrait pipelines, and anatomical completeness is a recognized gap this fills.