25万条部署轨迹变测试，VQS答案正确率94%

重点关注

• TraceDance: An Automated System for Building Agent Behavior Benchmarks from Real-World Agent Deployment Traces score 10
  入选 HF Daily Papers；HF 热度: 56 upvotes (+4)；有代码实现；关键词(3): deployment, tool use, coding
• CompoWorld: Compositional Environment Scaling for General Agents score 10
  入选 HF Daily Papers；HF 热度: 27 upvotes (+4)；有代码实现；关键词(3): scaling, fine-tuning, coding
• Program-Verified Self-Evolution for Vision-Language Models score 9
  入选 HF Daily Papers；HF 热度: 10 upvotes (+3)；有代码实现；关键词(1): vision-language
• YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality score 10
  入选 HF Daily Papers；HF 热度: 171 upvotes (+4)；有代码实现；关键词(1): agentic
• Diffusion Reward Models score 10
  入选 HF Daily Papers；HF 热度: 22 upvotes (+4)；有代码实现；关键词(2): lightweight, RLHF
• SpatialSpeak: QA-Native Reconstruction with Local and Global Context for Spatial Chain-of-Thought Reasoning score 9
  入选 HF Daily Papers；HF 热度: 18 upvotes (+3)；有代码实现；关键词(3): pretraining, reasoning, vision-language
• Beyond Timestamps: Decision-Aligned On-Policy Distillation for Long-Horizon Agents score 6
  入选 HF Daily Papers；有代码实现；关键词(2): distillation, GRPO

也值得关注

• Generalization Dynamics of LM Pre-training score 4
  机构: Google；关键词(2): pre-training, reasoning
• FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs score 4
  关键词(4): lightweight, deployment, fine-tuning, vision-language；顶会接收: NeurIPS
• Structured Sparse Memory for Recurrent Reasoning score 4
  关键词(2): reasoning, synthetic data；顶会接收: NeurIPS
• TT-VidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining score 12
  入选 HF Daily Papers；HF 热度: 10 upvotes (+3)；有代码实现；关键词(3): compression, fine-tuning, pretraining；顶会接收: NeurIPS
• SceneScaffold: Active Scene-State Construction for Unified 3D Scene Understanding score 4
  关键词(1): reasoning；顶会接收: NeurIPS
• Quizzing the Translation: A Prover-Grounded Evaluation Metric for NL$\rightarrow$FOL score 4
  关键词(1): reasoning；顶会接收: EMNLP
• Closing the Cross-Dialect Gap: Query Plans as a Portable Interface in Text-to-SQL score 4
  机构: ETH Zurich；关键词(2): production, fine-tuning
• DuraS2ST: Chain-of-Thought and Reinforcement Learning for Duration-Aligned Speech-to-Speech Translation score 4
  关键词(2): reasoning, open-source；顶会接收: EMNLP
• Counterfactual Rollout Replay: Forkable Environments as Free Process Rewards for Software Engineering Agents score 4
  关键词(1): GRPO；顶会接收: NeurIPS
• CAME: Company-Aware Evidence-Memory Experts for Interpretable Quarter-Ahead Revenue Forecasting score 3
  顶会接收: EMNLP
• The Price of Locality: Why Forward-Forward Underperforms Backpropagation? score 3
  顶会接收: NeurIPS
• Safe Score Matching: Diffusion Policies with Hamilton-Jacobi Reachability for Online Safe Reinforcement Learning score 3
  顶会接收: NeurIPS
• When Noise Meets Long-Tail: Feature-Threshold Dual Calibration for Robust Pseudo-Labeling score 3
  顶会接收: NeurIPS
• Geometric Inductive Biases for Semi-Supervised Equalization: The Constellation-Aware Transformer score 3
  顶会接收: NeurIPS
• Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction score 3
  顶会接收: NeurIPS