Position paper: 'Reasoning traces' is a dangerous label for LLM intermediate tokens

A position paper led by Subbarao Kambhampati pushes back on how the field talks about intermediate token generation — the practice of having a language model emit a stream of tokens before its final answer to boost performance on reasoning tasks. The authors object to the now-common habit of calling this output a ‘reasoning trace’ or ‘thinking trace,’ arguing that such language quietly imports a human metaphor: it implies the tokens mirror the deliberate steps a person takes to solve a hard problem, and that reading them offers an interpretable window into the model’s internal process.

The core claim is that this anthropomorphization is not a harmless shorthand. Treating intermediate tokens as genuine thought, the paper contends, misrepresents what the models actually do, encourages users to trust these traces as faithful explanations, and steers research toward questionable premises about interpretability and reliability. The authors marshal evidence that the traces do not reliably correspond to how models arrive at answers.

The piece is framed as a call to the community rather than an empirical result, urging researchers and practitioners to drop the loaded terminology. Its relevance has grown as ‘reasoning’ models have become a central marketing and product category, making the gap between the cognitive language used to describe them and their actual mechanics a practical concern for anyone relying on chain-of-thought output as an explanation of model behavior.