Test-time compute is the work a model or inference system uses to produce an answer, including generating tokens and evaluating candidates. It differs from computation spent training the model.
Snell and colleagues’ 2024 study examined verifier-guided search and adaptive response strategies. The most effective approach depended on prompt difficulty, so more computation alone does not promise a better answer.
Reasoning models are one way to use inference computation. Chain-of-thought prompting can elicit intermediate steps; it does not itself set an inference budget. Monte Carlo tree search is an older example from game playing, where running more simulations at decision time can improve the chosen move. Compare answer quality, elapsed time and cost for your workload before choosing a strategy. Episode 11 discusses DeepSeek, and the AI reasoning overview places these terms alongside the research.
Sources
- Snell et al.: Scaling LLM Test-Time Compute Optimally — Studies verifier-guided search and adaptive response strategies; effectiveness varies with prompt difficulty.
Go deeper
- Lightman et al.: Let's Verify Step by Step paper
Inspect how trained reward models rank multiple candidate math solutions in best-of-N search. Separate the training feedback from the inference-time selection step and check the study's task limits.