DoSA相较SGLang实现最高1.94倍首token加速

重点关注

• Agent-Reflex: Teaching Language Agents to Act Reflectively score 10
  入选 HF Daily Papers；HF 热度: 195 upvotes (+4)；有代码实现；关键词(5): scaling, distillation, fine-tuning, agentic, reasoning
• WorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution score 10
  入选 HF Daily Papers；HF 热度: 82 upvotes (+4)；有代码实现；关键词(2): scaling, GRPO
• Vela: User Behavior as Supervision Sparks LLM Intelligence score 10
  入选 HF Daily Papers；HF 热度: 77 upvotes (+4)；有代码实现；关键词(2): fine-tuning, synthetic data
• Learning How to Remember: Memory Modality Selection for GUI Agents score 10
  入选 HF Daily Papers；HF 热度: 41 upvotes (+4)；有代码实现；关键词(1): vision-language
• Self-Distillation with Verbal Regularization score 10
  入选 HF Daily Papers；HF 热度: 32 upvotes (+4)；有代码实现；关键词(2): distillation, reasoning
• Harness^2: Recursive Agent Harnessing for an Open World score 9
  入选 HF Daily Papers；HF 热度: 96 upvotes (+4)；有代码实现
• DoSA: Dots Sparse Attention score 9
  入选 HF Daily Papers；HF 热度: 73 upvotes (+4)；有代码实现
• JevGPT: Thinking Fast and Slow in Autonomous Research score 9
  入选 HF Daily Papers；HF 热度: 28 upvotes (+4)；有代码实现
• ClawLoop: Self-Improving Agents via Harness-Native Training Loop score 8
  入选 HF Daily Papers；HF 热度: 27 upvotes (+4)；关键词(2): deployment, post-training
• LifeSearch: Benchmarking Deep Research Agents for Real-World Life and Productivity score 7
  入选 HF Daily Papers；HF 热度: 16 upvotes (+3)；关键词(1): reasoning

也值得关注

• Training LLMs for Multi-Agent Collaboration with Verifiable Information Rewards score 4
  入选 HF Daily Papers；关键词(1): post-training
• GameSmith: Scaling Zero-Sum Self-Play with Synthesised Games score 4
  机构: Cambridge；关键词(4): scaling, distillation, post-training, reasoning
• Splicing Hints into Reasoning Traces: Learning from Demonstrator Failures score 4
  机构: Google；关键词(1): reasoning
• Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation score 4
  机构: Stanford；关键词(1): reasoning
• Scaling Laws for Multimodal Data Mixtures score 4
  机构: Microsoft；关键词(2): scaling, MoE
• DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models score 4
  机构: Carnegie Mellon；关键词(2): fine-tuning, pretraining
• CoverSynth: Coverage-Guided Generation of Retrieval Evaluation Sets score 4
  机构: Amazon；关键词(1): deployment
• Latent Taxonomies: How Property Induction Develops in Language Model Pre-training score 4
  机构: EPFL；关键词(2): pre-training, data curation
• Unfair and Unsafe: Everyday Bias Triggers Emergent Misalignment score 4
  机构: Harvard；关键词(2): fine-tuning, fine-tune
• Handoff Curves: Initial Selection, Later Handoff, and Supervision Targets in Multi-turn LLM Routing score 4
  机构: Huawei；关键词(1): deployment
• Game-Creation-Agents: Autonomous Game Production score 4
  机构: Tencent；关键词(1): production
• Size Generalization for Diffusion Routing Solvers score 4
  机构: NVIDIA；关键词(1): GRPO
• Releasing the Inner Jev: A Simple Recipe for Open Decision Models score 4
  机构: Microsoft Research；关键词(3): serving, fine-tuning, fine-tune
• SHELDON: Reasoning through Dual Search over Hypothesis and Observation Spaces score 4
  机构: Stanford；关键词(2): scaling, reasoning
• Code-Switching Detection via Language-Selective Dimensions in Speech Foundation Models score 4
  机构: Apple；关键词(2): lightweight, pretraining