Chain-of-Thought Prompting

By Satwik · February 26, 2026

Chain-of-thought prompting showed that simply asking a large model to work through a problem step by step, rather than jumping to an answer, dramatically improved performance on arithmetic, symbolic, and commonsense reasoning. The intervention is almost embarrassingly simple - a few worked examples that show intermediate steps, or even just the phrase "let's think step by step" - yet the gains were large and mostly appeared only in sufficiently large models.

The finding was important for two reasons. It suggested that a lot of latent capability in big models is unlocked by inference-time prompting rather than more training, and it tied cleanly into the emergence discussion, since the benefit of reasoning traces grew sharply with scale.

For our work there are practical angles. First, chain-of-thought is a capability amplifier available to anyone at inference time with no fine-tuning, which means a deployed model's demonstrated skill is a lower bound on its elicitable skill. Evaluations that do not use reasoning prompts can badly underestimate what a model can do, including on dangerous tasks. Second, the reasoning trace is an interpretability surface: it exposes a legible chain we can inspect, though the stated reasoning does not always faithfully reflect the true computation, so it can mislead as well as inform.

Chain-of-thought reframed prompting from a trick into a core elicitation method, and it is now a baseline consideration in any serious capability assessment.