12k样本赢金融SOTA，CUDA优化快35%

重点关注

• Agentic Planning with Reasoning for Image Styling via Offline RL score 8
  机构: Microsoft；入选 HF Daily Papers；HF 热度: 2 upvotes (+1)；关键词(4): post-training, agentic, reasoning, synthetic data
• HY-WU (Part I): An Extensible Functional Neural Memory Framework and An Instantiation in Text-Guided Image Editing score 8
  机构: Tencent；入选 HF Daily Papers；有代码实现
• Unlocking Data Value in Finance: A Study on Distillation and Difficulty-Aware Training score 7
  入选 HF Daily Papers；HF 热度: 11 upvotes (+3)；关键词(5): distillation, deployment, post-training, reasoning, open-source
• AutoResearch-RL: Perpetual Self-Evaluating Reinforcement Learning Agents for Autonomous Neural Architecture Discovery score 5
  入选 HF Daily Papers；HF 热度: 4 upvotes (+1)；关键词(2): PPO, pretraining
• MedSteer: Counterfactual Endoscopic Synthesis via Training-Free Activation Steering score 5
  入选 HF Daily Papers；有代码实现

也值得关注

• Making LLMs Optimize Multi-Scenario CUDA Kernels Like Experts score 4
  入选 HF Daily Papers；HF 热度: 2 upvotes (+1)
• PresentBench: A Fine-Grained Rubric-Based Benchmark for Slide Generation score 4
  入选 HF Daily Papers；关键词(1): deployment
• Self-Supervised Multi-Modal World Model with 4D Space-Time Embedding score 4
  机构: Mila；关键词(3): scaling, vision-language, open source
• VirtueBench: Evaluating Trustworthiness under Uncertainty in Long Video Understanding score 4
  关键词(2): vision-language, open-source；顶会接收: CVPR
• $\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving score 4
  关键词(2): fine-tuning, reasoning；顶会接收: ICLR
• Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbation score 4
  关键词(2): retrieval-augmented, RAG；顶会接收: ICLR
• Learning to Rank the Initial Branching Order of SAT Solvers score 3
  机构: Harvard
• A Distributed Gaussian Process Model for Multi-Robot Mapping score 3
  机构: Imperial College
• ConfHit: Conformal Generative Design with Oracle Free Guarantees score 3
  顶会接收: ICLR