Sign In

MO-CAPO: Multi-Objective Cost-Aware Prompt Optimization

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

์ €์ž

Jan Bussing, Moritz Schlager, Timo Hei{\ss}, Tom Zehle, Matthias Feurer

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ ์„ฑ๋Šฅ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ถ”๋ก  ๋น„์šฉ๊นŒ์ง€ ๊ณ ๋ คํ•˜๋Š” ๋‹ค๋ชฉ์  ํ”„๋กฌํ”„ํŠธ ์ตœ์ ํ™”์˜ ํ•„์š”์„ฑ์„ ์ œ๊ธฐํ•˜๋ฉฐ, ๊ธฐ์กด ๋ฐฉ๋ฒ•๋ก ์˜ ์„ฑ๋Šฅ ์ค‘์‹ฌ์  ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ ๋น„ํšจ์œจ์ ์ธ ์ตœ์ ํ™” ๋ฐฉ์‹์„ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด MO-CAPO๋ผ๋Š” ์ƒˆ๋กœ์šด ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. MO-CAPO๋Š” ์„ฑ๋Šฅ๊ณผ ์ถ”๋ก  ๋น„์šฉ์„ ๋™์‹œ์— ์ตœ์ ํ™”ํ•˜๊ณ  ์˜ˆ์‚ฐ ํ• ๋‹น์„ ํ†ตํ•ด ๋น„์šฉ ํšจ์œจ์ ์ธ ํƒ์ƒ‰์„ ์ˆ˜ํ–‰ํ•˜๋ฉฐ, LLM ์ถ”๋ก ์˜ ์ „์ฒด์ ์ธ ๊ณ„์‚ฐ ํ”„๋กœํŒŒ์ผ์„ ๋ฐ˜์˜ํ•˜๋Š” ๋ฐฐํฌ ์ง€ํ–ฅ์ ์ธ ๋น„์šฉ ๋ชฉํ‘œ๋ฅผ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
์„ฑ๋Šฅ-๋น„์šฉ ํŠธ๋ ˆ์ด๋“œ์˜คํ”„ ํƒ์ƒ‰: MO-CAPO๋Š” ๋‹ค์–‘ํ•œ ์„ฑ๋Šฅ-๋น„์šฉ ๊ท ํ˜•์ ์„ ๊ฐ€์ง„ ํ”„๋กฌํ”„ํŠธ ์ง‘ํ•ฉ์„ ํšจ์œจ์ ์œผ๋กœ ๋ฐœ๊ฒฌํ•˜์—ฌ, ๋‹จ์ผ ๋ชฉํ‘œ ์ตœ์ ํ™” ๋ฐฉ๋ฒ•์œผ๋กœ๋Š” ์ฐพ๊ธฐ ์–ด๋ ค์šด ์‹ค์šฉ์ ์ธ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
โ€ข
๋น„์šฉ ํšจ์œจ์ ์ธ ์ตœ์ ํ™”: ์˜ˆ์‚ฐ ํ• ๋‹น์„ ํ™œ์šฉํ•˜์—ฌ ๊ธฐ์กด NSGA-II ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•๋ณด๋‹ค ํ›จ์”ฌ ์ ์€ ์˜ˆ์‚ฐ์œผ๋กœ๋„ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋„์ถœํ•˜๋ฉฐ, ์ „์ฒด์ ์ธ ์ตœ์ ํ™” ํšจ์œจ์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค.
โ€ข
์‹ค์šฉ์ ์ธ ํ‰๊ฐ€ ์ง€ํ‘œ ๋„์ž…: ๋…ธ์ด์ฆˆ R2 ๋ฐ ๊ทผ์‚ฌ ์˜ค์ฐจ(approximation gap)์™€ ๊ฐ™์€ ์ง€ํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ผ๋ฐ˜ํ™” ๋ฐ ๊ฐ•๊ฑด์„ฑ(robustness)์„ ํ‰๊ฐ€ํ•จ์œผ๋กœ์จ, ๋ณด๋‹ค ํ˜„์‹ค์ ์ธ ์†”๋ฃจ์…˜ ํ’ˆ์งˆ ํ‰๊ฐ€๊ฐ€ ๊ฐ€๋Šฅํ•ด์กŒ์Šต๋‹ˆ๋‹ค.
โ€ข
ํ–ฅํ›„ ๊ณผ์ œ: ์‹ค์ œ LLM ๋ฐฐํฌ ํ™˜๊ฒฝ์—์„œ MO-CAPO์˜ ์žฅ๊ธฐ์ ์ธ ์„ฑ๋Šฅ๊ณผ ์•ˆ์ •์„ฑ์„ ์ถ”๊ฐ€์ ์œผ๋กœ ๊ฒ€์ฆํ•˜๊ณ , ๋”์šฑ ๋ณต์žกํ•˜๊ณ  ๋‹ค์–‘ํ•œ ์ œ์•ฝ ์กฐ๊ฑด(์˜ˆ: ์‘๋‹ต ์‹œ๊ฐ„, ์—๋„ˆ์ง€ ์†Œ๋น„)์„ ํ†ตํ•ฉํ•˜๋Š” ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๐Ÿ‘