Test-time compute
The idea of spending more compute at inference (not training) — let the model think longer, generate more candidate answers, and pick the best.
Research from DeepMind and OpenAI in 2024 showed that a smaller model with 10× more inference budget can beat a 10× larger model with standard decoding. Techniques: long chain-of-thought (o1, R1), best-of-N sampling, tree search (ToT), process reward models.
The shift changed LLM deployment economics — inference more expensive per query, but training amortised. The entire "reasoning model" paradigm is test-time compute scaling in practice.