Sign In

Decision Potential Surface: A Theoretical and Practical Approximation of Large Language Model Decision Boundary

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

์ €์ž

Zi Liang, Zhiyao Wu, Haoyang Shang, Yulin Jin, Qingqing Ye, Huadi Zheng, Peizhao Hu, Haibo Hu

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์˜ ์˜์‚ฌ๊ฒฐ์ • ๊ฒฝ๊ณ„๋ฅผ ๋ถ„์„ํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ธฐ์กด์˜ ๊ณ„์‚ฐ์  ๋น„ํ˜„์‹ค์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด '์˜์‚ฌ๊ฒฐ์ • ์ž ์žฌ๋ ฅ ํ‘œ๋ฉด(Decision Potential Surface, DPS)'์ด๋ผ๋Š” ์ƒˆ๋กœ์šด ๊ฐœ๋…์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. DPS๋Š” ๊ฐ ์ž…๋ ฅ์— ๋Œ€ํ•œ ํด๋ž˜์Šค ๊ตฌ๋ณ„ ์‹ ๋ขฐ๋„๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์˜์‚ฌ๊ฒฐ์ • ๊ฒฝ๊ณ„์˜ ์ž ์žฌ๋ ฅ์„ ํฌ์ฐฉํ•˜๋ฉฐ, ์ด๋ก ์ ์œผ๋กœ DPS์˜ ๋“ฑ๊ณ ์„ ์ด LLM์˜ ์˜์‚ฌ๊ฒฐ์ • ๊ฒฝ๊ณ„์™€ ๋™์ผํ•จ์„ ์ฆ๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ œํ•œ๋œ K๊ฐœ์˜ ์ƒ˜ํ”Œ๋งŒ์„ ์‚ฌ์šฉํ•˜์—ฌ LLM์˜ ์˜์‚ฌ๊ฒฐ์ • ๊ฒฝ๊ณ„๋ฅผ ๊ฑฐ์˜ ์˜ค์ฐจ ์—†์ด ๊ทผ์‚ฌํ•˜๋Š” ์‹ค์šฉ์ ์ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ธ K-DPS๋ฅผ ์ œ์‹œํ•˜๊ณ , ์ƒ˜ํ”Œ๋ง ํšŸ์ˆ˜์— ๋”ฐ๋ฅธ ์˜ค์ฐจ ๋ณ€ํ™”๋ฅผ ์ด๋ก ์ ์œผ๋กœ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
LLM์˜ ๋‚ด๋ถ€ ๋™์ž‘ ๋ฐ ์˜์‚ฌ๊ฒฐ์ • ๊ณผ์ •์„ ํ•ด์„ํ•˜๋Š” ์ƒˆ๋กœ์šด ์ด๋ก ์  ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
โ€ข
LLM์˜ ์˜์‚ฌ๊ฒฐ์ • ๊ฒฝ๊ณ„ ๋ถ„์„์— ๋Œ€ํ•œ ์‹ค์งˆ์ ์ธ ๊ณ„์‚ฐ ๊ฐ€๋Šฅ์„ฑ์„ ์—ด์–ด์ฃผ์–ด, ๋ชจ๋ธ์˜ ํŽธํ–ฅ์„ฑ ๋ถ„์„, ๋ณด์•ˆ ์ทจ์•ฝ์  ํƒ์ง€ ๋“ฑ์— ํ™œ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
โ€ข
K-DPS ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ LLM์˜ ์˜์‚ฌ๊ฒฐ์ • ๊ฒฝ๊ณ„๋ฅผ ํšจ์œจ์ ์œผ๋กœ ๊ทผ์‚ฌํ•  ์ˆ˜ ์žˆ๋Š” ์‹ค์šฉ์ ์ธ ๋„๊ตฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
โ€ข
๋ณธ ์—ฐ๊ตฌ์—์„œ ์ œ์‹œ๋œ DPS ๋ฐ K-DPS ๋ฐฉ๋ฒ•๋ก ์ด ์‹ค์ œ ๋‹ค์–‘ํ•œ LLM ์•„ํ‚คํ…์ฒ˜ ๋ฐ ๋ณต์žกํ•œ ์ž‘์—…์—์„œ ์–ผ๋งˆ๋‚˜ ํšจ๊ณผ์ ์ผ์ง€์— ๋Œ€ํ•œ ์ถ”๊ฐ€์ ์ธ ์‹ค์ฆ์  ๊ฒ€์ฆ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
๐Ÿ‘