Sign In

IH-Challenge: A Training Dataset to Improve Instruction Hierarchy on Frontier LLMs

Created by
  • Haebom
Category
Empty

์ €์ž

Chuan Guo (Michael Pokorny), Juan Felipe Ceron Uribe (Michael Pokorny), Sicheng Zhu (Michael Pokorny), Christopher A. Choquette-Choo (Michael Pokorny), Steph Lin (Michael Pokorny), Nikhil Kandpal (Michael Pokorny), Milad Nasr (Michael Pokorny), Rai (Michael Pokorny), Sam Toyer, Miles Wang, Yaodong Yu, Alex Beutel, Kai Xiao

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์ด ์‹œ์Šคํ…œ, ๊ฐœ๋ฐœ์ž, ์‚ฌ์šฉ์ž, ๋„๊ตฌ ์ง€์นจ ๊ฐ„์˜ ์ถฉ๋Œ ์‹œ ์šฐ์„ ์ˆœ์œ„๋ฅผ ๋ถ€์—ฌํ•˜๋Š” '์ง€์นจ ๊ณ„์ธต(Instruction Hierarchy, IH)'์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ์ƒˆ๋กœ์šด ํ•™์Šต ๋ฐ์ดํ„ฐ์…‹์ธ IH-Challenge๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. IH-Challenge๋ฅผ ํ™œ์šฉํ•œ ๊ฐ•ํ™” ํ•™์Šต์€ LLM์˜ IH ๊ฐ•๊ฑด์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๊ณ , ์•ˆ์ „ํ•˜์ง€ ์•Š์€ ํ–‰๋™์„ ์ค„์ด๋ฉฐ, ์ „๋ฐ˜์ ์ธ ์œ ์šฉ์„ฑ์„ ์œ ์ง€ํ•˜๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
LLM์˜ ์ง€์นจ ๊ณ„์ธต(IH) ๊ฐ•๊ฑด์„ฑ์„ ์ฒด๊ณ„์ ์œผ๋กœ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ์‹ค์งˆ์ ์ธ ํ•™์Šต ๋ฐฉ๋ฒ•๋ก ๊ณผ ๋ฐ์ดํ„ฐ์…‹์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
โ€ข
IH-Challenge๋ฅผ ํ†ตํ•œ ํ•™์Šต์€ LLM์˜ ์•ˆ์ „์„ฑ(์˜ˆ: ํƒˆ์˜ฅ, ํ”„๋กฌํ”„ํŠธ ์ฃผ์ž… ๋ฐฉ์ง€)์„ ํ–ฅ์ƒ์‹œํ‚ค๋ฉด์„œ๋„ ์œ ์šฉ์„ฑ์„ ์œ ์ง€ํ•˜๋Š” ๋ฐ ๊ธฐ์—ฌํ•ฉ๋‹ˆ๋‹ค.
โ€ข
์ œ์•ˆ๋œ ๋ฐฉ๋ฒ•๋ก ์€ ํ˜„์žฌ LLM์˜ ๋ณต์žกํ•œ ์ง€์นจ ์ถฉ๋Œ ์ƒํ™ฉ์„ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ํšจ๊ณผ์ ์ด์ง€๋งŒ, ์—ฌ์ „ํžˆ ์ƒˆ๋กœ์šด ์œ ํ˜•์˜ ๊ณต๊ฒฉ์— ๋Œ€ํ•œ ์™„์ „ํ•œ ๋ฐฉ์–ด๋ฅผ ๋ณด์žฅํ•˜๊ธฐ๋Š” ์–ด๋ ต์Šต๋‹ˆ๋‹ค.
๐Ÿ‘