Categories of Inference-Time Scaling for Improved LLM Reasoning
OpenAI
Inference-time scaling improves LLM answer quality by using more compute during inference. This article categorizes approaches and reviews recent papers, showing that major LLM providers rely on such techniques.
Inference-time scaling has become a key method to improve LLM performance by allocating extra compute at inference. The article groups techniques into categories like Chain-of-Thought prompting, Self-Consistency, Best-of-N ranking, Rejection Sampling with a Verifier, Self-Refinement, and Search Over Solution Paths. It also overviews recent papers and notes that proprietary LLMs use these methods. The author split a book chapter on this topic into two parts, available in early access, and shares additional ideas and notes that didn't fit the final chapter.
- Сокращения
- LLM = Large Language Model — большая языковая модель
Source: Sebastian Raschka —
original