LLM reasoning tokens are scratch pads, not thoughts

A paper arguing that builders and researchers should stop anthropomorphizing intermediate tokens in chain-of-thought models got significant discussion. The core argument: calling these tokens 'reasoning' or 'thinking' is wrong. They function more like scratch paper that helps the model compute a better output. Framing them as genuine reasoning assigns intentionality and internal state that isn't there.

The thread had a genuinely interesting counterpoint. Someone noted that the 'aha' moments models produce in their chain-of-thought are statistically useful even if they don't reflect any internal state change. The model is not experiencing insight. It is generating tokens that correlate with better downstream outputs. The distinction matters for how you build on top of these systems.

A related thread was building a tool called Vomit specifically to clean up Claude 5's verbose token output using a second LLM as an editor. That tool exists because anthropomorphizing the verbose output leads users to trust it more than they should. The output feels considered and thoughtful. It often isn't.