Sign In

ZeroUnlearn: Few-Shot Knowledge Unlearning in Large Language Models

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

์ €์ž

Yujie Lin, Chengyi Yang, Zhishang Xiang, Yiping Song, Jinsong Su

๐Ÿ’ก ๊ฐœ์š”

๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ๋ฐฉ๋Œ€ํ•œ ์›น ๋ฐ์ดํ„ฐ ํ•™์Šต์œผ๋กœ ์ธํ•ด ๋ฏผ๊ฐํ•œ ์ •๋ณด๋ฅผ ํฌํ•จํ•˜๊ฒŒ ๋˜์–ด ์‚ฌ์ƒํ™œ ๋ฐ ์•ˆ์ „ ๋ฌธ์ œ๊ฐ€ ์ œ๊ธฐ๋ฉ๋‹ˆ๋‹ค. ๋ณธ ์—ฐ๊ตฌ๋Š” ๊ธฐ์กด์˜ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ฑฐ๋‚˜ ์„ฑ๋Šฅ ์ €ํ•˜๋ฅผ ์œ ๋ฐœํ•˜๋Š” ์žฌํ•™์Šต ๋ฐ ๊ฐ•์ œ ํŒŒ์ธํŠœ๋‹ ๋ฐฉ์‹ ๋Œ€์‹ , ๋ชจ๋ธ ํŽธ์ง‘์„ ํ†ตํ•ด ์ง€์‹์„ ์žฌ๋งคํ•‘ํ•˜๋Š” ์ƒˆ๋กœ์šด ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ œ์•ˆ๋œ ZeroUnlearn๋Š” ์†Œ์ˆ˜์˜ ์˜ˆ์‹œ๋งŒ์œผ๋กœ ๋ฏผ๊ฐํ•œ ์ž…๋ ฅ์„ ์ค‘๋ฆฝ์ ์ธ ์ƒํƒœ๋กœ ๋ฎ์–ด์“ฐ๊ณ  ๊ธฐ์กด ํ‘œํ˜„์„ ์ œ๊ฑฐํ•˜์—ฌ ํšจ์œจ์ ์ด๊ณ  ์ •ํ™•ํ•œ ์ง€์‹ ์ œ๊ฑฐ๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
ํšจ์œจ์ ์ธ ์ง€์‹ ์ œ๊ฑฐ: ์ œ๋กœ์ƒท(Zero-shot) ๋˜๋Š” ํ“จ์ƒท(Few-shot) ๋ฐฉ์‹์œผ๋กœ ๋ชจ๋ธ ํŽธ์ง‘์„ ํ†ตํ•ด ๊ธฐ์กด ๋ฐฉ๋ฒ•๋ก  ๋Œ€๋น„ ํ›จ์”ฌ ์ ์€ ์—ฐ์‚ฐ๋Ÿ‰์œผ๋กœ ํŠน์ • ์ง€์‹์„ ์ œ๊ฑฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
โ€ข
๊ด€๋ จ ์ง€์‹ ๋ณด์กด: ๋ชจ๋ธ ํŽธ์ง‘ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ์ œ๊ฑฐํ•˜๋ ค๋Š” ์ง€์‹ ์™ธ์— ๋‹ค๋ฅธ ์œ ์šฉํ•œ ์ผ๋ฐ˜ ์ง€์‹์˜ ์„ฑ๋Šฅ ์ €ํ•˜๋ฅผ ์ตœ์†Œํ™”ํ•ฉ๋‹ˆ๋‹ค.
โ€ข
๋‹ค์ค‘ ์ƒ˜ํ”Œ ๋ฐ ์ผ๋ฐ˜ํ™”: ์ œ์•ˆ๋œ ๋ฐฉ๋ฒ•๋ก ์„ ํ™•์žฅํ•˜์—ฌ ์—ฌ๋Ÿฌ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ์ง€์‹ ์ œ๊ฑฐ๋ฅผ ์ง€์›ํ•˜๋ฉฐ, ํ–ฅํ›„ ๋” ๋ณต์žกํ•˜๊ณ  ๋‹ค์–‘ํ•œ ํ˜•ํƒœ์˜ ๋ฏผ๊ฐํ•œ ์ •๋ณด ์ œ๊ฑฐ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
๐Ÿ‘