Sign In

Memory-Guided Tree Search with Cross-Branch Knowledge Transfer for LLM Solver Synthesis

์ž‘์„ฑ์ž
  • Haebom
์นดํ…Œ๊ณ ๋ฆฌ
Empty

์ €์ž

Fatemeh Haji, Javier Delarosa Quiros, Peyman Najafirad

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ์—ฐ๊ตฌ๋Š” ์กฐํ•ฉ ์ตœ์ ํ™” ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์œ„ํ•œ LLM ๊ธฐ๋ฐ˜ ์†”๋ฒ„ ํ•ฉ์„ฑ์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๋ฐ˜๋ณต์ ์ธ ์ œ์•ฝ ์œ„๋ฐ˜ ๋ฐ ์œ ์‚ฌ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ˆ˜๋ ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด MEMOIR์ด๋ผ๋Š” ๋ฉ”๋ชจ๋ฆฌ ๊ฐ€์ด๋“œ ํŠธ๋ฆฌ ํƒ์ƒ‰ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•˜๋ฉฐ, ์ด๋Š” ๋ถ„๊ธฐ๋ณ„ ๊ตญ์†Œ ๋ฉ”๋ชจ๋ฆฌ์™€ ์ „์—ญ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ง€์‹ ์ „์†ก์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, MEMOIR์€ ๊ธฐ์กด ๋ฒ ์ด์Šค๋ผ์ธ ๋Œ€๋น„ ๋†’์€ ์†”๋ฃจ์…˜ ์œ ํšจ์„ฑ๊ณผ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•˜๋ฉฐ, ์ผ๊ด€์„ฑ ์žˆ๋Š” ๊ฐœ์„ ์„ ์ œ๊ณตํ•จ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
ํšจ๊ณผ์ ์ธ ์ง€์‹ ์ „์†ก ๋ฉ”์ปค๋‹ˆ์ฆ˜: MEMOIR์˜ ๋‘ ๊ณ„์ธต ๋ฉ”๋ชจ๋ฆฌ ๊ตฌ์กฐ๋Š” ์†”๋ฒ„ ํ•ฉ์„ฑ ๊ณผ์ •์—์„œ ๋ฐœ์ƒํ•˜๋Š” ์‹คํŒจ ํŒจํ„ด ๋ฐ ์„ฑ๊ณต์ ์ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ •๋ณด๋ฅผ ํšจ์œจ์ ์œผ๋กœ ์ €์žฅํ•˜๊ณ  ๊ณต์œ ํ•จ์œผ๋กœ์จ, ๋ฐ˜๋ณต์ ์ธ ์˜ค๋ฅ˜๋ฅผ ์ค„์ด๊ณ  ํƒ์ƒ‰ ํšจ์œจ์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค.
โ€ข
์•ˆ์ •์ ์ธ ์†”๋ฒ„ ์„ฑ๋Šฅ: ๋ฉ”๋ชจ๋ฆฌ ๊ฐ€์ด๋“œ ํƒ์ƒ‰์€ ๋ฌด์ž‘์œ„ ์ƒ˜ํ”Œ๋ง ๋ณ€๋™์„ฑ์— ์˜์กดํ•˜๋Š” ๊ธฐ์กด ๋ฐฉ๋ฒ•๊ณผ ๋‹ฌ๋ฆฌ, ์ผ๊ด€๋˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์†”๋ฒ„ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
โ€ข
๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ ๋ฐ ํ™•์žฅ์„ฑ: ๋ณต์žกํ•œ ๋ฌธ์ œ๋‚˜ ๋Œ€๊ทœ๋ชจ ํƒ์ƒ‰ ๊ณต๊ฐ„์—์„œ ๋ฉ”๋ชจ๋ฆฌ ๊ด€๋ฆฌ ์ „๋žต์ด ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ๊ณผ, ์ œ์•ˆ๋œ ๋ฉ”๋ชจ๋ฆฌ ๊ณ„์ธต ๊ตฌ์กฐ์˜ ํ™•์žฅ์„ฑ์— ๋Œ€ํ•œ ์ถ”๊ฐ€์ ์ธ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
๐Ÿ‘