Qwen3.5-4B以5%视觉token将9基准平均保留性能从68.6%升至82.3%

重点关注

• AdaTutoRank: Learning to Rerank Document Sets via Adaptive Tutoring Optimization for RAG and Deep Research score 9
  入选 HF Daily Papers；HF 热度: 13 upvotes (+3)；有代码实现；关键词(2): distillation, RAG
• Rethinking Training-Inference Mismatch in LLM Reinforcement Learning: Where It Arises and How to Correct It score 8
  入选 HF Daily Papers；HF 热度: 5 upvotes (+2)；有代码实现；关键词(1): reasoning
• Fewer Tokens, More Self-Teaching: On-Policy Self-Distillation for Extreme Visual Token Reduction score 7
  入选 HF Daily Papers；HF 热度: 2 upvotes (+1)；有代码实现；关键词(1): distillation
• MassAlloc Attention: Let Attention Allocate Its Own Compute score 6
  入选 HF Daily Papers；有代码实现；关键词(3): scaling, latency, reasoning

也值得关注

• CoWindow Attention: Full Causal Coverage Is a Collective Property score 4
  入选 HF Daily Papers；关键词(3): scaling, latency, reasoning
• Spectral Reversal: Counteracting Singular Value Bias for Graph Prompting score 4
  关键词(2): fine-tuning, pre-training；顶会接收: NeurIPS
• Noisy Test-Time Reinforcement Learning for Code LLMs score 4
  关键词(1): fine-tuning；顶会接收: EMNLP
• OptiArena: Can LLMs Improve Executable Algorithms under Fixed Resource Budgets? score 4
  关键词(1): coding；顶会接收: EMNLP
• Solving Every Step Is Not Enough: Milestone Oracles Reveal a Composition Gap in LLM Math Reasoning score 4
  关键词(2): code generation, reasoning；顶会接收: NeurIPS
• GLIDE: Generalized Layer-wise Intrinsic Distributional Evaluation for Heterogeneous LLM Agents score 4
  关键词(2): lightweight, reasoning；顶会接收: EMNLP
• Delayed Supervision for Test-Time Language Models score 4
  机构: MIT；关键词(1): post-training
• PULSE: Identifying Demonstration-Utility Features with Sparse Autoencoders score 4
  关键词(1): reasoning；顶会接收: NeurIPS
• Stabilizing the Dynamic Low-Rank Training score 4
  关键词(2): lightweight, compression；顶会接收: NeurIPS
• World SLAM Model: Joint World Modeling for SLAM and Navigation score 4
  机构: Tsinghua；关键词(1): embodied
• Business Compromise Detection with Agentic AI and LLM-driven Knowledge Discovery score 4
  关键词(2): edge, agentic；顶会接收: EMNLP
• LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations score 4
  关键词(3): fine-tuning, pretraining, jailbreak；顶会接收: NeurIPS
• Efficient Dynamic Algorithms for Graph Neural Networks with Non-Linear Propagation score 4
  关键词(1): edge；顶会接收: NeurIPS
• HyperLabel: Multi-Label Classification via Hypergraph-Based Label Correlation Modeling score 3
  机构: Amazon
• SCLATE: a Substrate for Continual-Learning Agent Training and Evaluation score 3
  机构: Apple