Agent跑久了上下文会「腐烂」，但好的上下文管理策略能大幅挽回性能

重点关注

• LOCA-bench: Benchmarking Language Agents Under Controllable and Extreme Context Growth score 9
  入选 HF Daily Papers；HF 热度: 23 upvotes (+4)；关键词(5): agent, agents, agentic, benchmark, open-source
• Thinking Makes LLM Agents Introverted: How Mandatory Thinking Can Backfire in User-Engaged Agents score 7
  入选 HF Daily Papers；HF 热度: 6 upvotes (+2)；关键词(3): agent, agents, reasoning
• Rolling Sink: Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusion score 6
  入选 HF Daily Papers；HF 热度: 6 upvotes (+2)；关键词(1): diffusion
• SoulX-Singer: Towards High-Quality Zero-Shot Singing Voice Synthesis score 6
  入选 HF Daily Papers；HF 热度: 2 upvotes (+1)；关键词(6): deployment, production, speech, benchmark, evaluation
• Learning-guided Kansa collocation for forward and inverse PDEs beyond linearity score 5
  入选 HF Daily Papers；关键词(2): benchmark, evaluation
• Pruning as a Cooperative Game: Surrogate-Assisted Layer Contribution Estimation for Large Language Models score 5
  关键词(6): efficient, lightweight, pruning, deployment, inference；顶会接收: ICLR
• Harpoon: Generalised Manifold Guidance for Conditional Tabular Diffusion score 5
  关键词(2): inference, diffusion；顶会接收: ICLR
• Rethinking Latency Denial-of-Service: Attacking the LLM Serving Framework, Not the Model score 5
  机构: Huawei；关键词(4): serving, inference, latency, cost
• FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging score 5
  关键词(4): efficient, compression, inference, attention；顶会接收: ICLR

也值得关注

• TerraBind: Fast and Accurate Binding Affinity Prediction through Coarse Structural Representations score 2
  关键词(5): fast, inference, diffusion, multimodal, benchmark
• Learnable Chernoff Baselines for Inference-Time Alignment score 2
  关键词(4): scaling, inference, alignment, diffusion
• HypRAG: Hyperbolic Dense Retrieval for Retrieval Augmented Generation score 2
  关键词(3): transformer, retrieval-augmented, RAG
• Riemannian MeanFlow score 2
  关键词(4): efficient, inference, diffusion, cost
• Geo-Code: A Code Framework for Reverse Code Generation from Geometric Images Based on Two-Stage Multi-Agent Evolution score 2
  关键词(4): agent, code generation, reasoning, multimodal
• Learning to Continually Learn via Meta-learning Agentic Memory Designs score 5
  入选 HF Daily Papers；关键词(4): efficient, agent, agentic, reasoning
• Preference Conditioned Multi-Objective Reinforcement Learning: Decomposed, Diversity-Driven Policy Optimization score 2
  关键词(2): PPO, preference
• PAND: Prompt-Aware Neighborhood Distillation for Lightweight Fine-Grained Visual Classification score 2
  关键词(4): lightweight, distillation, alignment, vision-language
• SRR-Judge: Step-Level Rating and Refinement for Enhancing Search-Integrated Reasoning in Search Agents score 2
  关键词(8): efficient, fine-tuning, post-training, agent, agents
• Generative Reasoning Re-ranker score 2
  关键词(3): fine-tuning, reasoning, recommendation
• Talk, Judge, Cooperate: Gossip-Driven Indirect Reciprocity in Self-Interested LLM Agents score 2
  关键词(3): agents, agentic, reasoning
• Attn-GS: Attention-Guided Context Compression for Efficient Personalized LLMs score 2
  关键词(6): efficient, compression, inference, latency, fine-tuning
• Still Manual? Automated Linter Configuration via DSL-Based LLM Compilation of Coding Standards score 2
  关键词(2): efficiency, coding
• Uncertainty-Aware Counterfactual Traffic Signal Control with Predictive Safety and Starvation-Avoidance Constraints Using Vision-Based Sensing score 2
  关键词(2): deployment, safety