“As a source: Position”
Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!
反对把中间 token 称作推理或思考痕迹的立场文章
这是一篇立场论文,指出语言模型在给出答案前生成的中间 token 常被称为“推理痕迹”甚至“思考痕迹”,这种拟人化暗示它们像人类解题步骤一样可读、可解释。作者给出证据,认为这种说法不是无害的比喻,会混淆人们对模型本质和用法的理解,并导致有问题的研究,呼吁社区不要再这样拟人化。
评论者在争论“思维链”是否真是推理时引用它,说明中间 token 可能缺乏语义含义、推理痕迹不能当作模型思考过程的证据。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“[1] Marvin Minsky via Rao Kampampathi and students: Position”
“It is quite likely that the intermediate tokens don’t have ‘semantic import’[0]”
“I thought it was pretty well understood amongst experts that trying to interpret so-called "chain of thought" is anthopomorphization and that the actual "thought process" of models is already opaque.”
“This is not a given, see Position and [LINK]”
“Still not meaningful -- Position; even for local models, the reasoning traces are often filtered and summarized to sound sensible to humans.”
- (How) Do reasoning models reason? ↗nyrikki · 2025-04-21 · 2 分
- (How) Do reasoning models reason? ↗YeGoblynQueenne · 2025-05-06 · 3 分
- Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces ↗felineflock · 2025-06-04 · 1 分
- Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Trac ↗xiaoyu2006 · 2026-06-26 · 1 分
- Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces ↗nunodonato · 2026-08-19 · 1 分