Test-Time Compute Lets AI Reason Harder Without Retraining on More Data

Traditionally, AI improvement has focused on scaling model size during training, a process known as training-time compute. Test-time compute is an alternative approach where a model performs extra reasoning steps at the moment a question is asked, rather than relying solely on what was learned during training. Techniques involved include chain-of-thought reasoning, generating multiple candidate answers, self-correction, and tree-based search over reasoning paths. However, the approach comes with trade-offs: slower response times, higher computational costs, and diminishing returns beyond a certain thinking threshold. Experts also caution that more reasoning time cannot fix a fundamentally flawed starting approach and works best on problems the model already has the knowledge to solve.