引用emp17344 · 2025-05
“This paper makes the claim that the “reasoning” output these models produce isn’t a necessary component of these models, and is likely unrelated to the actual final output:”
指出论文主张模型输出的推理并非必要组成部分,很可能与最终输出无关
用错误推理痕迹训练的模型表现与正确痕迹相当
作者从零训练 Transformer,让它们学习可形式化验证的推理痕迹及对应答案,发现即使在完全正确的痕迹上训练,模型也可能输出无效的推理痕迹却得到正确答案;在与题目毫无关系的被破坏痕迹上训练,效果与正确痕迹相近,甚至在分布外任务上泛化更好。GRPO 强化学习提高了答案准确率但并未提高痕迹有效性,痕迹长度也与问题的计算复杂度基本无关,论文据此提醒不要把中间 token 当作类人推理。
HN 上的人用它支持中间推理 token 未必有语义、模型不一定在真的推理这一观点。
“This paper makes the claim that the “reasoning” output these models produce isn’t a necessary component of these models, and is likely unrelated to the actual final output:”
“Subbarao Kambhampati and his team has several papers strongly suggesting that the intermediate tokens are of no symantic import.”
“If it was a reasoning machine then we would not see results like this:”
“"Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens" Beyond Semantics”