SFT收敛≠全学会，注意力劫持破防94%

也值得关注

• CoSToM:Causal-oriented Steering for Intrinsic Theory-of-Mind Alignment in Large Language Models score 4
  关键词(2): lightweight, reasoning；顶会接收: ACL
• Reason Only When Needed: Efficient Generative Reward Modeling via Model-Internal Uncertainty score 4
  关键词(2): lightweight, reasoning；顶会接收: ACL
• Learning Hierarchical and Geometry-Aware Graph Representations for Text-to-CAD score 4
  关键词(1): code generation；顶会接收: ICLR
• Why Supervised Fine-Tuning Fails to Learn: A Systematic Study of Incomplete Learning in Large Language Models score 4
  关键词(3): fine-tuning, post-training, pre-training；顶会接收: ACL
• Think in Sentences: Explicit Sentence Boundaries Enhance Language Model's Capabilities score 4
  关键词(2): fine-tuning, reasoning；顶会接收: ACL
• Credit-Budgeted ICPC-Style Coding: When Agents Must Pay for Every Decision score 4
  关键词(1): coding；顶会接收: ICLR
• Seeing No Evil: Blinding Large Vision-Language Models to Safety Instructions via Adversarial Attention Hijacking score 4
  关键词(1): vision-language；顶会接收: ACL
• Who Wrote This Line? Evaluating the Detection of LLM-Generated Classical Chinese Poetry score 3
  顶会接收: ACL