Reasoning Trace

The tokens a model generates on the way to an answer, treated as a first-class object rather than as output. Whether the trace is shown to the user, whether it is trained on, and whether it is faithful to what actually determined the answer are three separate questions, and conflating them is the most common mistake made about them.

Faithfulness is the uncomfortable one. A trace is generated by the same process as the answer, so it is a plausible account rather than a log — a model can produce a trace that reaches the right conclusion by reasoning that is not what determined its output, and studies that perturb the trace and watch the answer not move are the standard way of showing this. Treating a trace as an explanation is therefore a stronger claim than treating it as scratch space, and only the second is safe.

It is scratch space in a real computational sense: a fixed-depth network computing in one forward pass is bounded by that depth, and generating intermediate tokens lets it use its own output as memory across many passes. That is why chains help on problems requiring serial steps and do nothing on problems that do not — the mechanism is extra serial computation, not the model explaining itself, and the two are constantly confused.