Generating and exposing an AI model’s intermediate reasoning steps to improve transparency and complex problem solving.
Also known as: CoT, chain-of-thought prompting
In plain English
Think of it like a student showing their working in a maths problem: the AI writes out the steps it took to reach an answer. That makes the process easier to understand and check for mistakes.
In practice
In tools you buy or use, chain of thought shows up as step-by-step explanations or "workings" the model produces for multi-step tasks, helping trace and verify results. It can improve accuracy on planning, debugging and extraction tasks but raises token costs, latency and the need to review possibly incorrect intermediate steps.
Under the hood
Chain of thought is produced by prompting or fine-tuning models to emit intermediate tokens that represent reasoning steps (via few-shot demonstrations, rationale-labelled data or synthetic supervision). It relies on autoregressive decoding and larger context/scratchpad usage to surface explicit token-level traces, improving compositional reasoning while increasing token usage and sensitivity to prompt design and sampling temperature.
Example
Can we enable Chain of Thought prompting so the model shows its working on this risk assessment?