打乱音视频练推理，6B模型超GPT-4o

重点关注

• OmniJigsaw: Enhancing Omni-Modal Reasoning via Modality-Orchestrated Reordering score 9
  入选 HF Daily Papers；HF 热度: 16 upvotes (+3)；有代码实现；关键词(2): post-training, reasoning
• Lighting-grounded Video Generation with Renderer-based Agent Reasoning score 9
  入选 HF Daily Papers；HF 热度: 5 upvotes (+2)；关键词(3): lightweight, production, reasoning；顶会接收: CVPR
• SkillClaw: Let Skills Evolve Collectively with Agentic Evolver score 8
  入选 HF Daily Papers；HF 热度: 204 upvotes (+4)；关键词(2): deployment, agentic
• MegaStyle: Constructing Diverse and Scalable Style Dataset via Consistent Text-to-Image Style Mapping score 8
  入选 HF Daily Papers；HF 热度: 83 upvotes (+4)；关键词(3): fine-tune, text-to-image, data curation
• Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces score 8
  入选 HF Daily Papers；HF 热度: 8 upvotes (+2)；有代码实现；关键词(1): synthetic data
• Small Vision-Language Models are Smart Compressors for Long Video Understanding score 8
  入选 HF Daily Papers；HF 热度: 9 upvotes (+2)；有代码实现；关键词(4): scaling, compression, distillation, vision-language
• FIT: A Large-Scale Dataset for Fit-Aware Virtual Try-On score 6
  入选 HF Daily Papers；HF 热度: 14 upvotes (+3)
• Faithful GRPO: Improving Visual Spatial Reasoning in Multimodal Language Models via Constrained Policy Optimization score 5
  入选 HF Daily Papers；HF 热度: 4 upvotes (+1)；关键词(2): GRPO, reasoning
• Beyond Stochastic Exploration: What Makes Training Data Valuable for Agentic Search score 5
  入选 HF Daily Papers；HF 热度: 2 upvotes (+1)；关键词(2): agentic, reasoning
• PokeGym: A Visually-Driven Long-Horizon Benchmark for Vision-Language Models score 5
  入选 HF Daily Papers；HF 热度: 3 upvotes (+1)；关键词(4): deployment, reasoning, vision-language, embodied

也值得关注

• Direct Segmentation without Logits Optimization for Training-Free Open-Vocabulary Semantic Segmentation score 4
  关键词(1): vision-language；顶会接收: CVPR
• GRASS: Gradient-based Adaptive Layer-wise Importance Sampling for Memory-efficient Large Language Model Fine-tuning score 4
  关键词(2): throughput, fine-tuning；顶会接收: ACL
• More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration score 4
  关键词(2): scaling, reasoning；顶会接收: ICLR
• ReRec: Reasoning-Augmented LLM-based Recommendation Assistant via Reinforcement Fine-tuning score 4
  关键词(2): fine-tuning, reasoning；顶会接收: ACL
• DSCA: Dynamic Subspace Concept Alignment for Lifelong VLM Editing score 4
  关键词(2): instruction tuning, reasoning；顶会接收: CVPR
• SceneScribe-1M: A Large-Scale Video Dataset with Comprehensive Geometric and Semantic Annotations score 4
  关键词(1): text-to-video；顶会接收: CVPR
• Few-Shot Incremental 3D Object Detection in Dynamic Indoor Environments score 4
  关键词(2): vision-language, embodied；顶会接收: CVPR
• Revise: A Framework for Revising OCRed text in Practical Information Systems with Data Contamination Strategy score 4
  关键词(1): synthetic data；顶会接收: ACL
• Aligning Agents via Planning: A Benchmark for Trajectory-Level Reward Modeling score 4
  关键词(3): RLHF, agentic, reasoning；顶会接收: ACL
• MedVR: Annotation-Free Medical Visual Reasoning via Agentic Reinforcement Learning score 4
  关键词(4): deployment, agentic, reasoning, vision-language；顶会接收: ICLR
• Scaling-Aware Data Selection for End-to-End Autonomous Driving Systems score 4
  关键词(1): scaling；顶会接收: CVPR
• Meta-learning In-Context Enables Training-Free Cross Subject Brain Decoding score 4
  关键词(1): fine-tuning；顶会接收: CVPR
• Cross-Modal Emotion Transfer for Emotion Editing in Talking Face Video score 3
  顶会接收: CVPR
• PolicyLong: Towards On-Policy Context Extension score 3
  机构: Tencent