Chain-of-Thought Monitoring

Chain-of-thought (CoT) monitoring uses a reasoning model's available intermediate reasoning text as a signal for detecting behavior that its final answer or actions may not reveal. A human or another model can inspect that trace alongside tool calls and outputs to flag, for example, plans to evade a test, misuse a tool, or pursue a goal outside the user's request.

Monitoring is different from chain-of-thought prompting. Prompting asks a model to reason through a problem; monitoring evaluates evidence from a model's reasoning process. It is also narrower than agent trace and observability, which record actions, tool results, timing, and outcomes whether or not reasoning text is available.

A useful monitor needs a defined target behavior, access to the relevant observations, and evaluation against labeled cases. Teams should measure false positives and missed cases, compare reasoning-based signals with action-only signals, and verify consequential external outcomes directly. A reasoning trace is neither a complete account of the model's computation nor proof that it will act as written. Its usefulness can change with training, task design, and the amount of reasoning generated.

Many application APIs do not expose raw reasoning text, so this technique may be available mainly to model developers or deployments with specific access to internal traces. Where traces are retained, access controls and retention limits matter because they can contain sensitive context. Agent evaluation should test the monitor on realistic trajectories and keep independent checks on tool actions and final state.

The research paper Monitoring Monitorability evaluates when reasoning traces help detect behavior and where that signal remains incomplete.