Categories of Inference-Time Scaling for Better LLM Reasoning

Summary

Inference-time scaling allocates additional computation and generation time after a language model has been trained, with the aim of improving answer quality and accuracy. Sebastian Raschka explains that the approach has become common among major LLM providers and has attracted substantial academic research. The article organizes the field into categories including chain-of-thought prompting, self-consistency, best-of-N ranking, rejection sampling with a verifier, self-refinement, and search over solution paths. It is based partly on experiments conducted while preparing two chapters for the book Build a Reasoning Model (From Scratch); extensive hyperparameter tuning led to thousands of runs. In the author’s reported setup, the base model’s accuracy increased from about 15 percent to around 52 percent. The discussion focuses on training-free inference methods that spend more compute without changing model weights, while noting that stronger training and inference scaling can be used together.