SAS直连任务损失，AIM分类达96%

重点关注

• SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking score 10
  入选 HF Daily Papers；HF 热度: 59 upvotes (+4)；有代码实现；关键词(4): lightweight, post-training, agentic, reasoning

也值得关注

• AIM: A Privacy-Aware Interoperable Memory Framework for Multi-Agent Multi-User LLM Systems score 4
  机构: Stanford；关键词(1): agentic
• BlueLM-GUI Technical Report: A Real-Device-Centric Flywheel for Self-Improving Mobile GUI Agents score 4
  机构: Oxford；关键词(6): deployment, production, fine-tuning, pre-training, agentic
• Beyond ID Embeddings: Process-Grounded Language Modeling for Cognitive Diagnosis score 4
  机构: Huawei；关键词(1): MoE
• SCOPE-OPSD: Fisher-Conditioned Privileged Subspaces for On-Policy Self-Distillation score 4
  机构: Alibaba；关键词(1): distillation
• When Rubrics Fail: Hallucinations Reveal Blind Spots in Medical AI Evaluation score 4
  机构: EPFL；关键词(1): deployment
• Can We Trust LLM Judges: A Study of Capability-Dependent Biases and Multi-Judge Ensemble for Bias Calibration score 3
  机构: Stanford
• Quantifying Consonant Contributions to Word Intelligibility via Acoustic Masking score 3
  机构: Carnegie Mellon
• SWARM: A Multilingual Human-Annotated Dataset for Russian Propaganda Detection in Search Engine Results score 3
  机构: Oxford