Token级路由吞吐提升2.01—64.15倍

重点关注

• TokenRouter: Efficient Serving System for Token-Level LLM Routing score 13
  入选 HF Daily Papers；HF 热度: 123 upvotes (+4)；有代码实现；关键词(3): production, serving, throughput；顶会接收: NeurIPS
• SuperNav: An Agentic Navigation System for Any Task in Any Scene score 10
  入选 HF Daily Papers；HF 热度: 66 upvotes (+4)；有代码实现；关键词(4): deployment, fine-tuning, fine-tune, agentic
• AgentGarten: Code Worlds for Evolving Agents score 10
  入选 HF Daily Papers；HF 热度: 145 upvotes (+4)；有代码实现；关键词(1): real-time
• VibeEdit: Image Editing with Canvas Instructions score 9
  入选 HF Daily Papers；HF 热度: 15 upvotes (+3)；有代码实现；关键词(1): fine-tuning
• From Prompting to Composing: A Spatial Canvas Interface for Poster Generation score 8
  入选 HF Daily Papers；HF 热度: 8 upvotes (+2)；有代码实现；关键词(1): agentic
• Reasoning-Informed Visual Editing score 9
  入选 HF Daily Papers；HF 热度: 17 upvotes (+3)；有代码实现；关键词(3): agentic, reasoning, open-source
• Chaos in the Text: Revealing the Modality Preference in Mixed-Modality Retrievers score 7
  入选 HF Daily Papers；HF 热度: 5 upvotes (+2)；有代码实现
• Scaling to Tens of Thousands of Test-Time Iterations with Loop-Native Attention Residuals score 7
  入选 HF Daily Papers；HF 热度: 4 upvotes (+1)；有代码实现；关键词(2): scaling, reasoning
• SparseDecoding: Decoding-Aware Pruning for Accurate and Efficient LLM Inference score 8
  入选 HF Daily Papers；HF 热度: 23 upvotes (+4)；关键词(2): pruning, latency

也值得关注

• V-CoLA: Vision Token Compression with Linear Attention score 6
  入选 HF Daily Papers；HF 热度: 9 upvotes (+2)；关键词(2): compression, vision-language
• Diffusion Meta-Prompting and Steering for Generalizable Foundation Model Adaptation score 4
  关键词(1): text-to-image；顶会接收: NeurIPS
• DiscoVL: Unveiling Disentangled C ross-Modal Representation Learning via Orthogonal Adversarial Regularization for V ision-Language Models score 4
  关键词(1): vision-language；顶会接收: ECCV
• Higher-Order Action Supervision Makes A Strong Policy Class score 4
  关键词(3): lightweight, deployment, robotics；顶会接收: NeurIPS
• Breaking the Group Size Barrier: Parameter-Efficient Group Dance Generation with Chain-of-Dancers score 4
  关键词(1): lightweight；顶会接收: NeurIPS
• From Retrieval to Reconstruction: Constructing Evolvable Cognitive Memory for Long-Term Dialogue score 4
  关键词(5): serving, retrieval-augmented, RAG, agentic, reasoning；顶会接收: EMNLP
• CRISP: Fixing Flying Pixels in Latent LiDAR Generation via Diffusion Decoding score 4
  关键词(1): edge；顶会接收: NeurIPS
• SAGE: Sink-Aware Guided Emphasis for Visual Grounding in Vision-Language Decoders score 4
  关键词(2): lightweight, vision-language；顶会接收: EMNLP
• Neural Networks for Temporal Pattern Recognition and Dynamic Arm Gesture Speed Estimation for Robot Control score 4
  机构: FAIR；关键词(3): deployment, real-time, robotics
• Thinking Inertia: LLMs Keep Thinking When Told Not To score 4
  关键词(1): reasoning；顶会接收: NeurIPS
• Pose-Free Feed-Forward 3D Inpainting via Learnable Mask Attention and Support Token Refinement score 4
  关键词(1): reasoning；顶会接收: NeurIPS
• Is Memorization Context-Sensitive? Prefix-Based Extraction Beyond Isolated Prefixes score 4
  关键词(2): retrieval-augmented, RAG；顶会接收: EMNLP
• EgoVoice: Proactive Spoken Assistance from Egocentric Multimodal Streams score 4
  关键词(2): fine-tune, open-source；顶会接收: EMNLP
• Ambient Discrete Diffusion: Using the Wrong Data at the Right Time for Data Efficient Learning score 4
  关键词(1): finetuning；顶会接收: NeurIPS
• VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation score 4
  关键词(3): lightweight, deployment, vision-language；顶会接收: NeurIPS