Chain-of-Thought prompting elicits intermediate reasoning steps from a language model rather than a single direct answer, improving accuracy on multi-step probl
Chain-of-thought (CoT) has the model reason step-by-step before answering: elicited by prompting ('think through this first') or built into reasoning models that generate internal deliberation natively. Externalizing intermediate steps dramatically improves math, logic, and multi-hop accuracy, and modern 'thinking' models scale this with controllable reasoning budgets.
CoT unlocked reliable multi-step reasoning and became the basis of 2025–26's reasoning-model wave: systems that trade inference compute for accuracy on hard problems. For builders it's a core quality lever, and its traces aid debugging, though they're explanations to verify, not guaranteed windows into the model's true process.
On multi-step problems, math, planning, analysis, tricky extraction, where direct answers err. Skip it for simple lookups: reasoning costs tokens and latency without accuracy gains there.
Not necessarily: models can produce plausible rationales that diverge from their actual computation. Treat traces as useful structure and debugging signal, but verify conclusions independently when stakes are high.
Models post-trained to deliberate internally at length before answering, with controllable thinking budgets. They dominate hard-reasoning benchmarks and let builders price accuracy: more thinking tokens, better answers, higher cost.