Signal & Sense

What it is

Chain-of-thought is a model working a problem in written intermediate steps before committing to an answer. The reasoning trace is that visible step-by-step text — once a prompting trick, now trained in: reasoning models produce long traces by default.

How it works

A transformer spends roughly fixed compute per emitted token, so writing steps out buys more serial computation, and each step conditions on the ones before — one hard prediction becomes many easier ones. Reasoning models learn this via RL that rewards traces ending in verified answers.

A concrete example

Ask 'a 3.4L jug fills how many full 150mL glasses?' and a direct answer often fumbles the conversion. Prompted to reason, the model writes: 3.4L is 3400mL; 3400/150 is about 22.7; so 22 full glasses. Each line is a smaller prediction than the whole problem at once.

Why it matters

Every frontier reasoning model is chain-of-thought scaled up with RL and test-time compute. Hidden reasoning tokens are billed as output on major APIs, so hard questions cost more than their visible answers suggest — and traces are your main surface for debugging agents.