Test-time compute refers to the computational resources expended during inference rather than during training. It is the practice of allocating additional processing budget at query time so a model can reason through more steps, verify its own outputs, or search over candidate solutions before returning a final answer.

Think of it like a student who has already studied but is allowed extra scratch paper during the exam. The knowledge is fixed, but given more time to work through problems, the student produces better answers.

A concrete instance is OpenAI's o1 family, which uses chain-of-thought reasoning internally before producing output. The model generates a long internal reasoning trace, then surfaces only the conclusion. This shifts cost from training runs to per-query budgets, which researchers can tune: a hard problem gets more compute, a trivial lookup gets less.

Test-time compute connects directly to Inference, which is the forward pass of an LLM producing tokens. It also intersects with Chain of Thought, where explicit intermediate steps improve accuracy, and with Process Reward Models that score reasoning traces rather than final answers. Scaling test-time compute is sometimes contrasted with scaling training compute, and both interact with Foundation Model capabilities. Retrieval-Augmented Generation (RAG) is a separate strategy that extends what the model can access at inference time, but without necessarily increasing reasoning depth.

You will encounter this term when comparing reasoning-optimized models, evaluating API cost trade-offs, or reading research on scaling laws beyond training data.