ML//chain of thought

Chain of thought is a prompting technique (the most effective in prompt engineering) in which the model writes out its reasoning before giving the answer, and it is used on math, logic and multi-step reasoning problems; the classic trigger is the line *Let's think step by step*.

Chain of thought is a prompting technique (the most effective in prompt engineering) in which the model writes out its reasoning before giving the answer, and it is used on math, logic and multi-step reasoning problems; the classic trigger is the line Let's think step by step. Wei et al. (2022) showed that it dramatically improves math, logic and multi-step reasoning. It works because the intermediate tokens change the context before the answer, which triggers a distributional shift toward the reasoned conclusion region of latent space.

The written reasoning is an output like any other. A chain of thought that reads well need not be a faithful account of the computation that produced the answer; the model can reach a conclusion one way and justify it another. Studying what really happens inside is the work of mechanistic interpretability, and checking the result against tests or sources is what makes it trustworthy (verifier).

It is the simplest form of test-time compute: more tokens spent on one request, and the weights unchanged afterwards.

Tree-of-thought generalizes it to a branching exploration of several reasoning paths.

Extended thinking is the trained, scaled version: the same principle (more reasoning tokens give better answers), optimized with RL and dedicated compute instead of a prompt hack.

Reasoning models (o1, o3, R1) prove CoT scales: trained thinking chains outperform prompted ones by orders of magnitude on hard problems.