Chain-of-Thought Monitoring

Using the model’s visible reasoning process as a supervision signal — a place to look for signs the model is pursuing something other than the task. Only useful to the extent the reasoning is a faithful account of what is happening.

That qualification is the whole difficulty. Faithfulness tests found models often use a hint without disclosing it, so the trace can be incomplete or misleading. It remains worth monitoring, as long as it is treated as a useful signal rather than a reliable transcript.

The practical posture is to read it and to verify independently. The trace is frequently the fastest route to understanding an unexpected action; it is also generated text, and the faithfulness tests show it can omit the actual cause of an answer.

Treating the reasoning trace as an audit log. It is generated output, and generated output can be incomplete in ways that look complete.

• blogReasoning models don't always say what they think ↗Anthropic Research · 2025

  Tests whether the visible chain of thought reflects the real cause of an answer. Often it does not, which caps how much you can trust a reasoning trace as an audit log.

• paperChain-of-thought prompting elicits reasoning in large language models ↗Wei et al. (arXiv) · 2022

  Few-shot demonstrations containing intermediate reasoning steps improved results on arithmetic, commonsense and symbolic reasoning tasks.