文档Agent导航≈碰运气，预填充加速1.82×

重点关注

• Trust Your Critic: Robust Reward Modeling and Reinforcement Learning for Faithful Image Editing and Generation score 13
  机构: Shanghai AI Lab；入选 HF Daily Papers；HF 热度: 22 upvotes (+4)；有代码实现；关键词(2): text-to-image, data curation
• EVATok: Adaptive Length Video Tokenization for Efficient Visual Autoregressive Generation score 12
  入选 HF Daily Papers；HF 热度: 10 upvotes (+3)；有代码实现；关键词(1): lightweight；顶会接收: CVPR
• EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models score 11
  机构: Shanghai AI Lab；入选 HF Daily Papers；HF 热度: 9 upvotes (+2)；有代码实现；关键词(2): scaling, reasoning
• Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections score 10
  入选 HF Daily Papers；HF 热度: 49 upvotes (+4)；有代码实现；关键词(2): agentic, reasoning
• IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse score 10
  入选 HF Daily Papers；HF 热度: 35 upvotes (+4)；有代码实现；关键词(5): lightweight, distillation, production, serving, agentic
• ShotVerse: Advancing Cinematic Camera Control for Text-Driven Multi-Shot Video Creation score 10
  入选 HF Daily Papers；HF 热度: 28 upvotes (+4)；有代码实现；关键词(1): vision-language
• XSkill: Continual Learning from Experience and Skills in Multimodal Agents score 10
  入选 HF Daily Papers；HF 热度: 20 upvotes (+4)；有代码实现；关键词(2): tool use, reasoning
• Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights score 10
  机构: Peking University；入选 HF Daily Papers；HF 热度: 3 upvotes (+1)；有代码实现；关键词(4): PPO, GRPO, post-training, pretraining
• Spatial-TTT: Streaming Visual-based Spatial Intelligence with Test-Time Training score 9
  入选 HF Daily Papers；HF 热度: 69 upvotes (+4)；有代码实现
• DVD: Deterministic Video Depth Estimation with Generative Priors score 9
  入选 HF Daily Papers；HF 热度: 17 upvotes (+3)；有代码实现；关键词(1): open-source

也值得关注

• Multi-Task Reinforcement Learning for Enhanced Multimodal LLM-as-a-Judge score 4
  入选 HF Daily Papers；HF 热度: 3 upvotes (+1)
• Deployment-Time Reliability of Learned Robot Policies score 4
  机构: Stanford；关键词(1): deployment
• Stay in your Lane: Role Specific Queries with Overlap Suppression Loss for Dense Video Captioning score 4
  关键词(1): lightweight；顶会接收: CVPR
• SPEGC: Continual Test-Time Adaptation via Semantic-Prompt-Enhanced Graph Clustering for Medical Image Segmentation score 4
  关键词(2): deployment, edge；顶会接收: CVPR
• LaMoGen: Language to Motion Generation Through LLM-Guided Symbolic Inference score 4
  关键词(1): reasoning；顶会接收: CVPR
• Tokenization Allows Multimodal Large Language Models to Understand, Generate and Edit Architectural Floor Plans score 4
  关键词(2): instruction tuning, reasoning；顶会接收: CVPR
• Stable Spike: Dual Consistency Optimization via Bitwise AND Operations for Spiking Neural Networks score 4
  关键词(1): latency；顶会接收: CVPR
• UCAN: Unified Convolutional Attention Network for Expansive Receptive Fields in Lightweight Super-Resolution score 4
  关键词(4): scaling, lightweight, distillation, deployment；顶会接收: CVPR
• Language Generation with Replay: A Learning-Theoretic View of Model Collapse score 4
  机构: INRIA；关键词(1): scaling
• Intrinsic Concept Extraction Based on Compositional Interpretability score 4
  关键词(1): text-to-image；顶会接收: CVPR
• AdaFuse: Accelerating Dynamic Adapter Inference via Token-Level Pre-Gating and Fused Kernel Optimization score 4
  关键词(3): latency, MoE, open-source；顶会接收: AAAI
• BTZSC: A Benchmark for Zero-Shot Text Classification Across Cross-Encoders, Embedding Models, Rerankers and LLMs score 4
  关键词(3): scaling, latency, fine-tuning；顶会接收: ICLR
• Cross-Domain Policy Optimization via Bellman Consistency and Hybrid Critics score 4
  关键词(1): state space；顶会接收: ICLR
• HATS: Hardness-Aware Trajectory Synthesis for GUI Agents score 4
  关键词(1): vision-language；顶会接收: CVPR
• IsoCompute Playbook: Optimally Scaling Sampling Compute for LLM RL score 4
  机构: Carnegie Mellon；关键词(3): scaling, post-training, pre-training