Pivot-SD在LLaDA-8B-Instruct上以200题、每题4次采样胜过全序列SFT，Queen七轮从1782升至2697

重点关注

• 4DCodeBench: Benchmarking Agents on Inverse Graphics of Dynamic Scenes score 12
  机构: Stanford；入选 HF Daily Papers；HF 热度: 12 upvotes (+3)；有代码实现；关键词(1): code generation
• Pivot-SD: Efficient Self-Distillation for Masked Diffusion Language Models score 11
  机构: University of Toronto；入选 HF Daily Papers；HF 热度: 38 upvotes (+4)；关键词(3): distillation, post-training, reasoning
• Multilingual GSM-Symbolic: What determines capability transfer across languages? score 10
  入选 HF Daily Papers；HF 热度: 29 upvotes (+4)；有代码实现；关键词(1): reasoning
• World Embedding Benchmark score 10
  入选 HF Daily Papers；HF 热度: 28 upvotes (+4)；有代码实现；关键词(2): lightweight, retrieval-augmented
• ProAR: Learning Prospective Reasoning with Autoregressive Video Models score 9
  入选 HF Daily Papers；HF 热度: 18 upvotes (+3)；有代码实现；关键词(3): lightweight, reasoning, embodied
• Language Models that Play Chess and Explain Their Moves score 9
  入选 HF Daily Papers；HF 热度: 15 upvotes (+3)；有代码实现；关键词(3): distillation, reasoning, robotics
• Collective Bias Mitigation via Model Routing and Collaboration score 9
  机构: MIT；入选 HF Daily Papers；HF 热度: 10 upvotes (+3)
• Native Action-Prior Learning from Videos for World Action Models score 8
  入选 HF Daily Papers；HF 热度: 62 upvotes (+4)；关键词(1): pretraining
• FrugalEvo: Towards Cost-Aware LLM-Guided Program Evolution score 8
  入选 HF Daily Papers；HF 热度: 15 upvotes (+3)；有代码实现
• HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents score 7
  入选 HF Daily Papers；HF 热度: 43 upvotes (+4)

也值得关注

• Follow the Winners: Conservative Policy Improvement with the Cross-Entropy Method for Critic-Free RFT score 4
  机构: Cambridge；关键词(6): fine-tuning, DPO, PPO, GRPO, post-training
• Geometry Meets Physics: Data-Efficient Pre-Training for Unstructured Neural PDE Solvers score 4
  关键词(2): fine-tuning, pre-training；顶会接收: NeurIPS
• Measure Less, Know More: Self-Supervised Test-Time Feature Acquisition score 4
  关键词(1): deployment；顶会接收: NeurIPS
• Recursive Harness Self-Improvement for Frontier Reasoning Data Synthesis score 4
  机构: Cambridge；关键词(3): GRPO, coding, reasoning
• Cephalonauts One: A deep fMRI dataset for decoding naturalistic speech in the human brain score 4
  关键词(1): scaling；顶会接收: NeurIPS
• What Should World Models Forget? Stratified Retention for Continual Adaptation score 4
  关键词(2): latency, reasoning；顶会接收: NeurIPS
• ZeroMAG: Zero-Shot Multimodal Adapter Generation for Plug-and-Play EEG Foundation Models score 3
  机构: Zhejiang University
• Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis score 3
  顶会接收: NeurIPS