Mistral自研TTS、扩散LLM加速4.7倍

重点关注

• Intern-S1-Pro: Scientific Multimodal Foundation Model at Trillion Scale score 8
  入选 HF Daily Papers；HF 热度: 100 upvotes (+4)；关键词(3): scaling, reasoning, open-source
• Voxtral TTS score 8
  入选 HF Daily Papers；HF 热度: 27 upvotes (+4)；关键词(1): quantization
• RealRestorer: Towards Generalizable Real-World Image Restoration with Large-Scale Image Editing Models score 10
  入选 HF Daily Papers；HF 热度: 43 upvotes (+4)；有代码实现；关键词(1): open-source
• MACRO: Advancing Multi-Reference Image Generation with Structured Long-Context Data score 10
  入选 HF Daily Papers；HF 热度: 26 upvotes (+4)；有代码实现；关键词(2): fine-tuning, reasoning
• PixelSmile: Toward Fine-Grained Facial Expression Editing score 9
  入选 HF Daily Papers；HF 热度: 105 upvotes (+4)；有代码实现
• MuRF: Unlocking the Multi-Scale Potential of Vision Foundation Models score 8
  入选 HF Daily Papers；HF 热度: 8 upvotes (+2)；有代码实现；关键词(1): serving
• S2D2: Fast Decoding for Diffusion LLMs via Training-Free Self-Speculation score 7
  入选 HF Daily Papers；HF 热度: 4 upvotes (+1)；有代码实现；关键词(1): lightweight
• Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes score 7
  入选 HF Daily Papers；HF 热度: 3 upvotes (+1)；有代码实现；关键词(4): distillation, post-training, agentic, reasoning

也值得关注

• Once-for-All Channel Mixers (HYPERTINYPW): Generative Compression for TinyML score 4
  机构: MIT；关键词(3): compression, quantization, latency
• MoE-GRPO: Optimizing Mixture-of-Experts via Reinforcement Learning in Vision-Language Models score 4
  关键词(3): GRPO, MoE, vision-language；顶会接收: CVPR
• MSRL: Scaling Generative Multimodal Reward Modeling via Multi-Stage Reinforcement Learning score 4
  关键词(3): scaling, distillation, reasoning；顶会接收: CVPR
• Photon: Speedup Volume Understanding with Efficient Multimodal Large Language Models score 4
  关键词(2): scaling, compression；顶会接收: ICLR
• Offline Decision Transformers for Neural Combinatorial Optimization: Surpassing Heuristics on the Traveling Salesman Problem score 4
  关键词(1): deployment；顶会接收: NeurIPS
• EagleNet: Energy-Aware Fine-Grained Relationship Learning Network for Text-Video Retrieval score 4
  关键词(1): vision-language；顶会接收: CVPR
• SliderQuant: Accurate Post-Training Quantization for LLMs score 4
  关键词(4): quantization, post-training, MoE, reasoning；顶会接收: ICLR
• Separate Before You Compress: The WWHO Tokenization Architecture score 4
  机构: OpenAI；关键词(2): compression, reasoning
• Adaptive Learned Image Compression with Graph Neural Networks score 4
  关键词(1): compression；顶会接收: CVPR
• Multimodal Dataset Distillation via Phased Teacher Models score 4
  关键词(2): compression, distillation；顶会接收: ICLR
• Beyond the Golden Data: Resolving the Motion-Vision Quality Dilemma via Timestep Selective Training score 4
  关键词(1): data curation；顶会接收: CVPR