THINK DEEPLY: WHAT IS TEST-TIME COMPUTE?Test-time compute (TTC) is the allocation of additional computation during inference to enable a large language model (LLM) to perform internal and deliberate reasoning to solve complex tasks ("think deeply before answering").Driven by this mechanism, current LLMs have shifted from immediate heuristic pattern completion to slower multistep reasoning.Nowadays, these new LLMs are often called "reasoning models."Unlike traditional language models constrained to single-pass autoregressive next-token predictions based on statistical regularities, they can explore, evaluate, and refine candidate solution paths before answering 1.In the nascent stage, Snell et al. 2 demonstrated that TTC can outperform simply scaling training-time resources, such as model parameters, pretraining data, or total training
Kim et al. (Thu,) studied this question.