Additional computation spent during inference (e.g., generating and evaluating many candidate answers) to improve output quality without retraining.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
While 'inference' covers basic forward-pass execution and 'scaling-law' covers training-time scaling, neither explains the distinct idea that spending more compute *at inference time* (chain-of-thought, search, self-consistency sampling) yields quality gains. This is the mechanism behind claim 1.4 and bridge B3, and it sits between the existing inference and scaling-law concepts without being reducible to either.