Stop When Further Reasoning Won't Help: Attention-State Adaptive Generation in Reasoning Models
当进一步推理无益时停止:推理模型中的注意力状态自适应生成
机构 * University of Electronic Science and Technology of China(电子科技大学) ; Ubiquitous Intelligence and Trusted Services Key Laboratory of Sichuan Province(四川省 ubiquitous 智能与可信服务重点实验室)
专题命中 测试时计算 :reasoning(title,abstract);CoT(abstract,abstract_cn);chain-of-thought(abstract);test-time compute(abstract)
AI总结 针对大型推理模型过度思考导致冗余和准确率下降的问题,提出无需训练的注意力状态自适应生成方法ASAG,通过推断推理状态动态调整生成策略,在多个基准上平均准确率提升3.2%,生成token减少近40%。
Comments ICML 2026 Spotlight