Capabilities & Reasoning

chain-of-thought reasoning

Chain-of-thought is the simple trick of having the model write out its intermediate steps before the final answer, the way a teacher asks you to show your working. Instead of jumping straight to forty-two, the model first says there were three baskets, each held fourteen apples, three times fourteen is forty-two. Often you get this just by adding a phrase like let us think step by step, and accuracy on arithmetic, logic, and word problems climbs sharply with no change to the model itself.

Why it helps is intuitive once you remember the model reasons by continuing text. A long jump from question to answer asks it to do all the work invisibly in one prediction; writing the steps turns that one hard prediction into a sequence of easy ones, each leaning on the line above. The intermediate tokens act as an external scratchpad the model can read back. The catch is that the written chain is a story the model tells, not a transcript of hidden computation, so it can sound rigorous while quietly being wrong.

P(\text{answer} \mid \text{question}) \;<\; P(\text{answer} \mid \text{question}, \text{steps})

Spelling out the steps makes the correct ending a more likely continuation than answering in one leap.

Also called
CoTstep-by-step reasoning