Back to the index
09/ 55

USING MODELS

Chain of thought.

CoT · intermediate reasoning

Intermediate reasoning steps generated as part of working toward an answer.

In plain words

A chain of thought is a sequence of intermediate steps that can help a language model work through a problem. The phrase can describe a prompting technique, a generated rationale, or internal reasoning tokens that a product may not display.

A closer look

Breaking a problem into steps can provide useful intermediate results for subsequent computation. A model might work out quantities, compare alternatives, and check a constraint before giving an answer. Training can also encourage models to use longer reasoning sequences for difficult tasks.

A written explanation is not necessarily a faithful record of the computation that produced an answer. Models can omit influential factors, rationalize errors, or produce a convincing explanation for an incorrect conclusion. Evaluate the answer and its verifiable steps rather than relying on how thoughtful the prose sounds.

In practice

AN EXAMPLE

For a packing problem, intermediate work can calculate the volume of each object before checking whether the total fits the box. The arithmetic and assumptions can be checked independently.

A useful distinction

More reasoning text does not automatically mean better reasoning. Long chains can introduce extra errors, and an explanation of a decision is not proof of humanlike thought or access to every internal process.

Sources & further reading

Wei et al. — Chain-of-Thought Prompting Elicits Reasoning (opens in a new tab)