Self-Discover: Large Language Models Self-Compose Reasoning StructuresWe introduce SELF-DISCOVER, a general framework for LLMs to self-discover the task-intrinsic reasoning structures to tackle complex reasoning problems that are challenging for typical prompting methods. Core to the framework is a self-discovery process where LLMs select multiple atomic reasoning modules such as critical thinking and step-by-step thinking, and compose them into an explicit reasoning structure for LLMs to follow during decoding. SELF-DISCOVER substantially improves GPT-4 and PaLM 2's performance on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, by as much as 32% compared to Chain of Thought (CoT). Furthermore, SELF-DISCOVER outperforms inference-intensive methods such as CoT-Self-Consistency by more than 20%, while requiring 10-40x fewer inference compute. Finally, we show that the self-discovered reasoning structures are universally applicable across model families: from PaLM 2-L to GPT-4, and from GPT-4 to Llama2, and share commonalities with human reasoning patterns.arxiv.org

์ด ๋…ผ๋ฌธ์€ SELF-DISCOVER๋ผ๋Š” ์ƒˆ๋กœ์šด ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(Large Language Models, LLMs)์ด ๋ณต์žกํ•œ ์ถ”๋ก  ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋„๋ก ํƒœ์Šคํฌ ๊ณ ์œ ์˜ ์ถ”๋ก  ๊ตฌ์กฐ๋ฅผ ์Šค์Šค๋กœ ๋ฐœ๊ฒฌํ•˜๊ฒŒ ํ•˜๋Š” ๊ณผ์ •์ž…๋‹ˆ๋‹ค. SELF-DISCOVER๋Š” ๋น„ํŒ์  ์‚ฌ๊ณ ์™€ ๋‹จ๊ณ„๋ณ„ ์‚ฌ๊ณ  ๊ฐ™์€ ์—ฌ๋Ÿฌ ์›์ž์  ์ถ”๋ก  ๋ชจ๋“ˆ์„ ์„ ํƒํ•˜๊ณ , ์ด๋ฅผ ๋ช…์‹œ์ ์ธ ์ถ”๋ก  ๊ตฌ์กฐ๋กœ ๊ตฌ์„ฑํ•˜์—ฌ ๋””์ฝ”๋”ฉํ•˜๋Š” ๋™์•ˆ ๋”ฐ๋ฅด๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ GPT-4์™€ PaLM 2์˜ ์„ฑ๋Šฅ์„ ์ƒ๋‹นํžˆ ํ–ฅ์ƒ์‹œํ‚ค๋ฉฐ, ํŠนํžˆ BigBench-Hard, ์ง€์ƒ ์—์ด์ „ํŠธ ์ถ”๋ก , MATH ๊ฐ™์€ ๋„์ „์ ์ธ ์ถ”๋ก  ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ธฐ์กด ๋ฐฉ์‹๋ณด๋‹ค ์ตœ๋Œ€ 32% ๊ฐœ์„ ๋ฉ๋‹ˆ๋‹คย