How do reasoning models perform intermediate computations during inference?

How do reasoning models perform intermediate computations during inference?

Inference is the process of using a trained model to respond to an input. Every neural network performs intermediate numerical computations during inference. Reasoning models are designed or trained to make additional computation useful for tasks that benefit from several dependent steps.

Numerical representations carry information

Inside an LLM, layers transform vectors into new vectors. These temporary activations encode information used to predict later tokens. They are not a simple collection of human-readable sentences.

When a model generates additional intermediate tokens, those tokens can become context for subsequent computation. This creates more opportunities to develop or revise a solution before producing the user-facing answer.

Systems may combine several mechanisms

For a small programming task, a system might generate a candidate solution, call a code tool, inspect the test result, and revise the code. Tool execution is a separate operation performed by the surrounding application.

Other tasks may be handled entirely within model inference. The exact design varies by system, and additional computation need not always appear as visible explanatory text.

Explanations are useful but incomplete evidence

OpenAI's reasoning guide distinguishes reasoning tokens from visible output tokens. The visible answer therefore need not reveal all computation used to produce it.

Even a displayed rationale should not be assumed to be a faithful transcript of the model's internal process. Evaluate the result with task-specific checks, such as running code or verifying a derivation, and request a concise explanation suitable for review.