Self-Discover: Large Language Models Self-Compose Reasoning Structures

We introduce SELF-DISCOVER, a general framework for LLMs to self-discover thetask-intrinsic reasoning structures to tackle complex reasoning problems thatare challenging for typical prompting methods. Core to the framework is aself-discovery process where LLMs select multiple atomic reasoning modules suchas critical thinking and step-by-step thinking, and compose them into anexplicit reasoning structure for LLMs to follow during decoding. SELF-DISCOVERsubstantially improves GPT-4 and PaLM 2’s performance on challenging reasoningbenchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, by asmuch as 32

Further reading#

• Access Paper in arXiv.org