Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces (2025)

	Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces (2025)(arxiv.org)
	285 points by nunodonato 2 days ago \| 239 comments
	tl;dr: This ICML 2026 position paper argues against calling the intermediate tokens produced by LLMs "reasoning traces" or "thinking traces," contending that such anthropomorphization is misleading rather than a harmless metaphor. The authors claim these labels misrepresent what the models actually do, encourage users to treat the traces as interpretable windows into model cognition, and drive questionable research directions. They call on the community to abandon this framing.
	HN Discussion: ↓Anthropomorphization is a harmless metaphor and not a real problem in the field ↑Intermediate tokens shouldn't be treated as interpretable; focus on reproducibility of computation instead ↑Anthropomorphizing LLMs is harmful and thinking traces should be treated as black boxes ↓Selective outrage; we've long used anthropomorphic terms for machines without harm ~RL training makes outputs functionally resemble human thinking, so the label has some basis