Daily AI Papers — July 29, 2026

1. HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone

Authors: Yuteng Wei, Jinming Ma, Jiawei Wang, Weitao Zhou, et al. (Simple AI)

Summary: Introduces a portable UMI (universal manipulation interface) data-capture rig with head-mounted stereo-inertial SLAM and dual wide-angle cameras that reaches 3mm end-effector accuracy without external tracking. Policies post-trained solely on this “robot-free” data match real-robot teleoperation performance across three VLA/world-action-model backbones, removing the need for a real-robot data “anchor.”

arXiv: arxiv.org/abs/2607.25895 Sources: HuggingFace Daily Papers Why trending: Top-voted paper of the day (110 upvotes) — tackles the costly bottleneck of scaling robot manipulation data and open-sources a 2,000-hour dataset (HiFi-UMI-2K).

2. A New Role for Relevance: Guiding Corpus Interaction in Agentic Search

Authors: Jiangnan Li, Yuqing Li, Mo Yu, Jinchao Zhang, Jie Zhou

Summary: Argues that document-level relevance alone can’t localize or verify evidence for complex questions, and proposes using relevance signals to guide fine-grained grep-style corpus exploration (“Direct Corpus Interaction”) instead of just top-k filtering. This narrows search faster and surfaces useful evidence earlier in agentic retrieval pipelines.

arXiv: arxiv.org/abs/2607.24223 Sources: HuggingFace Daily Papers Why trending: Second-highest engagement (78 upvotes) — directly relevant to the fast-growing agentic-search/RAG space.

3. ReDesign: Recovering Editable Design Structures from Images via Agentic Decomposition

Authors: Jooyeol Yun, Jintae Park, Hyesu Lim, Junha Hyung, Hyungjin Chung, Jaegul Choo

Summary: Presents an agentic framework that reconstructs editable design files (typography, vector geometry, colors, layering) from a raster image by composing specialized tools with a “graceful verification” step that prevents error accumulation. Introduces the Figma Edit Replay Benchmark (909 files, ~14.8K edit instructions) and beats layered-decomposition baselines on editability.

arXiv: arxiv.org/abs/2607.25565 Sources: HuggingFace Daily Papers Why trending: 53 upvotes — strong practical appeal for design tooling/agentic workflows.

4. Keep It InMind: Benchmarking the Implicit-Association Blind Spot in Agent Memory

Authors: Ruizhe Li, Mingxuan Du, Benfeng Xu, Zhendong Mao

Summary: Identifies a failure mode in long-term memory systems where retrieval assumes a stored memory textually resembles the query needing it (e.g., a tree-nut allergy note won’t surface for a macaron request despite the almond-flour link). Introduces a benchmark to measure this “implicit-association blind spot” in agent memory architectures.

arXiv: arxiv.org/abs/2607.24368 Sources: HuggingFace Daily Papers Why trending: 23 upvotes — surfaces a subtle, underexplored gap in the fast-growing agent-memory subfield.

5. CodeNib: A Multi-View Data System for Serving Repository Context to Coding Agents

Authors: Zhongming Yu, Hengjia Yu, Boqin Yuan, Shuting Zhao, et al.

Summary: Builds reusable lexical, dense, and structural indexes per repository commit to serve ranked search, symbol navigation, and bounded context to coding agents through a single runtime, avoiding disconnected/rebuilt indexes. Shows 8.7x–25.4x faster incremental updates and 50–87% fewer trajectory tokens versus grep/read baselines across five models.

arXiv: arxiv.org/abs/2607.25431 Sources: HuggingFace Daily Papers Why trending: 22 upvotes — coding-agent infrastructure is a hot topic alongside the related Agent Retrieval Bench paper (#19) released the same window.

6. Pass the Baton: Trajectory-Relayed On-Policy Distillation

Authors: Haolei Xu, Xiaowen Xu, Haiwen Hong, Zixuan Ni, Hongxing Li, Yiwen Qiu, Weiming Lu, Yongliang Shen

Summary: Diagnoses “prefix failure” in on-policy distillation, where a student that commits to a wrong reasoning path keeps building on that error. Proposes Relay-OPD, which lets the teacher briefly take over at detected failure points before handing control back to the student, improving Qwen3 students by up to +5.73% on math benchmarks while cutting training trajectory length over 50%.

arXiv: arxiv.org/abs/2607.26057 Sources: HuggingFace Daily Papers Why trending: 21 upvotes — practical, cheap improvement to a widely-used distillation recipe for reasoning models.

7. Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model

Authors: Senqiao Yang, Kaichen Zhang, Zhaoyang Jia, Jinghao Guo, et al.

Summary: Targets “Moravec’s paradox” in VLMs — strong offline visual reasoning but weak/inefficient streaming perception — with a codec-native tokenizer (Mage-ViT) that selectively encodes motion/entropy-rich regions, cutting visual tokens over 75%. The 4B model matches Qwen3-VL-4B on static tasks while beating a 15B baseline on video/spatial reasoning with up to 3.5x inference speedup.

arXiv: arxiv.org/abs/2607.24904 Sources: HuggingFace Daily Papers Why trending: 21 upvotes — efficient real-time multimodal streaming is a key direction for on-device/agentic vision systems.

8. Novel Claim or Déjà Vu? Rethinking “Contamination-Free” Dynamic Evaluation for Multimodal Automated Fact-Checking

Authors: Haorui He, Xinwen Chen, Dacheng Wen, Reynold Cheng, Francis C. M. Lau, Yupeng Li

Summary: Shows that static fact-checking benchmarks risk contamination since claims are often verifiable from an LLM’s internal knowledge, inflating scores. Proposes rethinking “dynamic” benchmark construction so claims genuinely require external, up-to-date evidence to test multimodal automated fact-checking systems fairly.

arXiv: arxiv.org/abs/2607.23514 Sources: HuggingFace Daily Papers Why trending: 12 upvotes — timely critique of benchmark validity as fact-checking systems get deployed more widely.

9. Shieldstral

Authors: Antonia Calvi, Avinash Sooriyarachchi, Giada Pistilli, Guillaume Lample, Maarten Buyl, et al. (Mistral)

Summary: A 3B-parameter policy-adaptive multimodal safety classifier that reformulates content moderation as a single binary yes/no task, unifying heterogeneous moderation taxonomies under one training framework built from ~54.1M samples. It matches or beats models nearly 7x its size on text safety and sets a new SOTA on multimodal safety classification.

arXiv: arxiv.org/abs/2607.25857 Sources: HuggingFace Daily Papers Why trending: 10 upvotes — from Mistral’s safety team (Guillaume Lample among authors), notable for small-model efficiency in trust & safety.

10. Wonder: Video World Model Done Better

Authors: Jiacong Xu, Hanwen Jiang, Zhixin Shu, Kalyan Sunkavalli, Vishal M. Patel, Yiqun Mei

Summary: Presents a real-time, camera-controllable video world model that lets users interactively explore a “playable world” generated from an image or video, with a dense coordinate-field camera conditioning scheme and sparse-attention memory for long-horizon coherence. Synthesizes diverse minute-scale videos at 16 FPS while preserving geometry and dynamics across long rollouts.

arXiv: arxiv.org/abs/2607.26037 Sources: HuggingFace Daily Papers Why trending: 9 upvotes — video world models for interactive exploration are a hot generative-video research direction.

11. PerceptionBench: Evaluating Atomic Visual Perception in Multimodal Large Language Models

Authors: Zichao Lin, Yifeng Xie, Bowen Qu, Haiming Wang, et al. (large multi-org author list)

Summary: Builds a bottom-up benchmark isolating ten “atomic” visual perception capabilities from reasoning/knowledge, using 3,000 verified short-answer questions derived from diagnosing failure points across 42 existing benchmarks. Finds no frontier MLLM exceeds 60% accuracy and that perception-related hallucination is the weakest capability across the board.

arXiv: arxiv.org/abs/2607.24957 Sources: HuggingFace Daily Papers Why trending: 9 upvotes — a sobering, well-constructed benchmark showing basic visual perception is still unsolved in top MLLMs.

12. MODUS: Decoder-Only Any-to-Any Modeling of Diverse Modalities

Authors: Mingqiao Ye, Zhaochong An, Zhitong Gao, Xian Liu, François Fleuret, et al.

Summary: Investigates decoder-only any-to-any multimodal modeling that treats every modality symmetrically as both input and output within one network, letting the model reuse strong pre-trained decoder-only priors instead of training encoder-decoder/diffusion architectures from scratch. Supports chained generation through intermediate modalities and self-verification by cross-checking its own outputs.

arXiv: arxiv.org/abs/2607.25948 Sources: HuggingFace Daily Papers Why trending: 8 upvotes — pushes any-to-any multimodal architecture toward reusing decoder-only LLM priors, a meaningfully different design choice.

13. Visual prompt engineering for video models

Authors: Robert Geirhos, Yuxuan Li, Thaddäus Wiedemer, Neha Kalibhat, Zi Wang, Mani Malek, Oyvind Tafjord, Kevin Swersky, Been Kim, Priyank Jaini

Summary: Asks whether video foundation models benefit from “visual prompt engineering” — automatically modifying the input image/scene rather than the text prompt — and finds it improves video reasoning performance, sometimes more than classic text prompt engineering or test-time scaling. Demonstrated on visual physics reasoning tasks where converting sketch-like scenes to photorealistic ones boosts accuracy.

arXiv: arxiv.org/abs/2607.25537 Sources: HuggingFace Daily Papers Why trending: 7 upvotes — notable author list (Been Kim, Robert Geirhos) and a fresh, low-cost technique as video models become reasoning foundation models.

14. Parallel Decoding Distillation for Fast Image and Video Generation

Authors: Neta Shaul, Chao Liu, Arash Vahdat, Julius Berner

Summary: Introduces Parallel Decoding Distillation (PDD), a trajectory-based distillation method that predicts multiple denoising steps per network evaluation without the instability of variational score distillation or adversarial losses. Achieves SOTA 4–8 NFE generation on LTX-2.3, Wan 14B, and Qwen-Image while notably preserving video diversity where VSD-based methods suffer mode collapse.

arXiv: arxiv.org/abs/2607.26004 Sources: HuggingFace Daily Papers Why trending: 6 upvotes — practical few-step generation speedup (NVIDIA authors) addressing a known diversity-collapse weakness in prior distillation methods.

15. OmniDelta: Skill-Driven Budget Allocation for Token Compression in OmniLLMs

Authors: Haoyang Huang, Wenjie Huang, Tianqi Xu, Hongyaoxing Gu, et al.

Summary: Tackles token-budget allocation (not just token selection) for compressing long audio-video sequences in omni-modal LLMs, showing raw query-to-media similarity is unreliable for allocating budget across modalities. The training-free method reallocates budget using skill pools and local complexity/redundancy, cutting GPU memory 22% and achieving 1.64x speedup at 25% token retention.

arXiv: arxiv.org/abs/2607.25669 Sources: HuggingFace Daily Papers Why trending: 5 upvotes — addresses a practical inference-cost bottleneck for omni-modal models as audio-video LLMs proliferate.

16. Where Quality Breaks in Compressed Short-Text Generation: Staged Bottleneck Localization

Authors: Alexey Gavrilov, Alan-Barsag Gazzaev, Sergey Muravyov

Summary: Studies whether quality loss in compressed-latent text generators comes from the codec discarding information or from the latent generator producing weak codes, using a controlled TinyStories case study with a hierarchical VQ-VAE-2 codec and masked discrete diffusion generator. Proposes a staged validation protocol to separate these failure modes so researchers spend compute fixing the right component.

arXiv: arxiv.org/abs/2607.24176 Sources: HuggingFace Daily Papers Why trending: 3 upvotes — a focused diagnostic methodology relevant to anyone building latent-space text generators.

17. Towards Robust Reinforcement Learning for Small-Scale Language Model Agents

Authors: Md Rezwanul Haque, Md. Milon Islam, Fakhri Karray

Summary: Systematically investigates why PPO-based RL alignment is unstable for small (70–500M parameter) language models, identifying three concrete failure modes: silent LoRA freezing, bfloat16 overflow in importance ratios, and reward-model-driven policy collapse. Proposes fixes (adapter merge-reinitialize, float32 PPO updates, reward whitening/importance guarding/weight rollback) that stabilize training and beat instruction-tuned baselines with less data.

arXiv: arxiv.org/abs/2607.25091 Sources: HuggingFace Daily Papers Why trending: 3 upvotes — actionable, reproducible engineering fixes for small-model RL, useful for resource-constrained fine-tuning.

18. Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion

Authors: Cédric Bonhomme, Alexandre Dulaunoy

Summary: Builds a reproducible classifier mapping CVE descriptions directly to MITRE ATT&CK techniques, trained on a curated 1,207-CVE gold set, roughly doubling recall@5 versus zero-shot embedding similarity. Critically, a careful replication study shows LLM-assisted label expansion provides no reliable improvement and can hurt rare-technique coverage — the earlier apparent gain was an evaluation artifact from noisy checkpoint selection.

arXiv: arxiv.org/abs/2607.25572 Sources: HuggingFace Daily Papers Why trending: 2 upvotes — a rigorous negative result on LLM-assisted data labeling, relevant caution for anyone using LLMs to expand training sets.

19. Agent Retrieval Bench: Evaluating Repository Context Retrieval for Coding Agents

Authors: Bowen Qin, Yi Xie

Summary: Introduces a 427-sample, 25-repository benchmark isolating the upstream context-acquisition stage of coding agents (finding the right files before patch generation), covering four positive-retrieval tasks plus a no-gold/counterfactual selective-retrieval subset. Finds no single retrieval method dominates across tasks and that logged agent trajectories miss every gold file on 27–35% of samples, exposing a real bottleneck upstream of patch quality.

arXiv: arxiv.org/abs/2607.24882 Sources: HuggingFace Daily Papers Why trending: 2 upvotes — complements CodeNib (#5) in probing coding-agent context retrieval, a fast-growing evaluation gap.

20. VisualPatchWorld: Code World Models as Latent Structured Representations for Planning

Authors: Jiaxin Bai, Jiaxuan Xiong

Summary: Proposes representing world dynamics as inspectable, editable code rather than opaque neural weights or hand-built physics engines, selecting a qualitative dynamical form via active probes and fitting its parameters from state-action traces. Achieves 69.0% mean planning success, beating the strongest code-based world-model baseline by 23.5 points, with the largest gains where choosing correct dynamics form matters most.

arXiv: arxiv.org/abs/2607.25236 Sources: HuggingFace Daily Papers Why trending: 1 upvote — a distinctive “world-model-as-code” angle bridging neural and symbolic approaches to planning.