Test-time compute allocates more computation during inference so a model can do additional reasoning before answering. It reflects a shift from relying on pretraining scale alone toward organizing computation for tasks that require reasoning. section 7.0
Ways to add reasoning at inference include:
- Chain-of-thought prompting decomposes a query into subtasks and gives the model an internal scratchpad, without changing its architecture or parameters. section 7.5.3
- A model can generate several reasoning chains, search among them, and select the answer that appears most consistent—a vote among its candidate thoughts. section 7.5.3
Extra deliberation can strengthen performance, but prompting is a limited lever when a task requires recomputation under novel constraints; it cannot substitute for mechanisms that search or enforce invariants. section 7.5.3 section 7.5
