Lottie动画直接生成，DPO自带防遗忘

重点关注

• MMR-Life: Piecing Together Real-life Scenes for Multimodal Multi-image Reasoning score 13
  入选 HF Daily Papers；HF 热度: 40 upvotes (+4)；有代码实现；关键词(1): reasoning；顶会接收: ICLR
• OmniLottie: Generating Vector Animations via Parameterized Lottie Tokens score 12
  入选 HF Daily Papers；HF 热度: 122 upvotes (+4)；有代码实现；顶会接收: CVPR
• RubricBench: Aligning Model-Generated Rubrics with Human Standards score 12
  机构: Microsoft；入选 HF Daily Papers；HF 热度: 48 upvotes (+4)；有代码实现
• Beyond Length Scaling: Synergizing Breadth and Depth for Generative Reward Models score 11
  机构: Microsoft；入选 HF Daily Papers；HF 热度: 29 upvotes (+4)；关键词(4): scaling, fine-tuning, reasoning, open-source
• OpenAutoNLU: Open Source AutoML Library for NLU score 10
  入选 HF Daily Papers；HF 热度: 41 upvotes (+4)；有代码实现；关键词(2): open-source, open source
• PhotoBench: Beyond Visual Matching Towards Personalized Intent-Driven Photo Retrieval score 10
  入选 HF Daily Papers；HF 热度: 20 upvotes (+4)；有代码实现；关键词(2): agentic, reasoning
• CoVe: Training Interactive Tool-Use Agents via Constraint-Guided Verification score 8
  入选 HF Daily Papers；HF 热度: 20 upvotes (+4)；关键词(3): fine-tuning, post-training, open-source
• WorldStereo: Bridging Camera-Guided Video Generation and Scene Reconstruction via 3D Geometric Memories score 8
  入选 HF Daily Papers；HF 热度: 14 upvotes (+3)；有代码实现
• Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance score 8
  入选 HF Daily Papers；HF 热度: 14 upvotes (+3)；有代码实现
• Surgical Post-Training: Cutting Errors, Keeping Knowledge score 8
  入选 HF Daily Papers；HF 热度: 9 upvotes (+2)；有代码实现；关键词(3): DPO, post-training, reasoning

也值得关注

• DynaMoE: Dynamic Token-Level Expert Activation with Layer-Wise Adaptive Capacity for Mixture-of-Experts Neural Networks score 4
  入选 HF Daily Papers；关键词(2): scaling, MoE
• Provable and Practical In-Context Policy Optimization for Self-Improvement score 4
  关键词(3): scaling, pretraining, reasoning；顶会接收: ICLR
• ASTRA-bench: Evaluating Tool-Use Agent Reasoning and Action Planning with Personal User Context score 4
  机构: Apple；关键词(1): reasoning
• MixerCSeg: An Efficient Mixer Architecture for Crack Segmentation via Decoupled Mamba Attention score 4
  关键词(2): edge, mamba；顶会接收: CVPR
• Continuous Exposure-Time Modeling for Realistic Atmospheric Turbulence Synthesis score 4
  关键词(1): synthetic data；顶会接收: CVPR
• UniTalking: A Unified Audio-Video Framework for Talking Portrait Generation score 4
  关键词(1): open-source；顶会接收: CVPR
• Unifying Language-Action Understanding and Generation for Autonomous Driving score 4
  机构: Zhejiang University；关键词(2): latency, vision-language
• CARE: Towards Clinical Accountability in Multi-Modal Medical Reasoning with an Evidence-Grounded Agentic Framework score 4
  关键词(2): agentic, reasoning；顶会接收: ICLR
• Learning to Draft: Adaptive Speculative Decoding with Reinforcement Learning score 4
  机构: Peking University；关键词(1): throughput
• FastLightGen: Fast and Light Video Generation with Fewer Steps and Parameters score 4
  关键词(3): lightweight, pruning, deployment；顶会接收: CVPR
• Cross-modal Identity Mapping: Minimizing Information Loss in Modality Conversion via Reinforcement Learning score 4
  关键词(3): fine-tuning, reasoning, vision-language；顶会接收: CVPR
• Action-Guided Attention for Video Action Anticipation score 4
  关键词(1): post-training；顶会接收: ICLR
• Meta-Learning Hyperparameters for Parameter Efficient Fine-Tuning score 4
  关键词(2): scaling, fine-tuning；顶会接收: CVPR
• Modular Memory is the Key to Continual Learning Agents score 4
  机构: Mila；关键词(1): pretraining
• AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations score 4
  关键词(2): RAG, agentic；顶会接收: ICLR