Test-Time Compute Scaling

A theoretical framework for scaling computation during the inference phase of large language models. By allocating additional computational resources at inference time—through mechanisms such as chain-of-thought reasoning and multi-step verification—output quality can be improved without modifying model parameters.