Categories of Inference-Time Scaling for Improved LLM Reasoning

Sebastian Raschka categorizes the landscape of inference-time scaling techniques for LLMs — methods that trade more compute at inference for better outputs — drawing from thousands of experimental runs. The piece synthesizes recent academic literature into clearer groupings and is excerpted from a new book chapter that improved a base model from ~15% to ~52% accuracy on a reasoning benchmark. Highly substantive reference for practitioners thinking about how to get more out of deployed models.