Sign In

XL-SafetyBench: A Country-Grounded Cross-Cultural Benchmark for LLM Safety and Cultural Sensitivity

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

์ €์ž

Dasol Choi, Eugenia Kim, Jaewon Noh, Sang Seo, Eunmi Kim, Myunggyo Oh, Yunjin Park, Brigitta Jesica Kartono, Josef Pichlmeier, Helena Berndt, Sai Krishna Mendu, Glenn Johannes Tungka, Ozlem Gok\c{c}e, Suresh Gehlot, Katherine Pratt, Amanda Minnich, Haon Park

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ ๊ธฐ์กด ์˜์–ด ์ค‘์‹ฌ์˜ LLM ์•ˆ์ „์„ฑ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ๊ตญ๊ฐ€๋ณ„ ํŠน์ˆ˜ํ•œ ํ”ผํ•ด๋ฅผ ๊ฐ„๊ณผํ•˜๊ณ  ๋ฌธํ™”์  ๋ฏผ๊ฐ์„ฑ์„ ๋ณดํŽธ์  ํ”ผํ•ด์™€ ๊ตฌ๋ถ„ํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด XL-SafetyBench๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. XL-SafetyBench๋Š” 10๊ฐœ ๊ตญ๊ฐ€-์–ธ์–ด ์Œ์— ๊ฑธ์ณ 5,500๊ฐœ์˜ ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค๋ฅผ ํฌํ•จํ•˜๋ฉฐ, ๊ตญ๊ฐ€๋ณ„ ํŠน์„ฑ์— ๊ธฐ๋ฐ˜ํ•œ ์ ๋Œ€์  ํ”„๋กฌํ”„ํŠธ์™€ ๋ฌธํ™”์  ๋ฏผ๊ฐ์„ฑ์„ ๋‚ดํฌํ•œ ์š”์ฒญ์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” LLM์˜ ํƒˆ์˜ฅ ๊ณต๊ฒฉ ๋ฐฉ์–ด ๋Šฅ๋ ฅ๊ณผ ๋ฌธํ™”์  ๋ฏผ๊ฐ์„ฑ ์ธ์ง€ ๋Šฅ๋ ฅ์„ ๋™์‹œ์— ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
๊ธฐ์กด ๋ฒค์น˜๋งˆํฌ์˜ ์˜์–ด ์ค‘์‹ฌ์„ฑ๊ณผ ๋ฒˆ์—ญ ๋ฐฉ์‹์˜ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ณ , ๊ตญ๊ฐ€๋ณ„ ๋งž์ถคํ˜• ์•ˆ์ „์„ฑ ํ‰๊ฐ€์˜ ์ค‘์š”์„ฑ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
โ€ข
LLM์˜ ์•ˆ์ „์„ฑ ํ‰๊ฐ€์— ์žˆ์–ด ํƒˆ์˜ฅ ๋ฐฉ์–ด ๋Šฅ๋ ฅ๊ณผ ๋ฌธํ™”์  ๋ฏผ๊ฐ์„ฑ ์ธ์ง€ ๋Šฅ๋ ฅ์ด ๋…๋ฆฝ์ ์œผ๋กœ ํ‰๊ฐ€๋˜์–ด์•ผ ํ•จ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
โ€ข
์ง€์—ญ LLM์˜ ์•ˆ์ „์„ฑ ์ง€ํ‘œ๊ฐ€ ์‹ค์ œ ์ •๋ ฌ๋ณด๋‹ค๋Š” ์ƒ์„ฑ ์‹คํŒจ์— ๊ธฐ์ธํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ์‹œ์‚ฌํ•˜๋ฉฐ, ์ง„์ •ํ•œ ์•ˆ์ „์„ฑ ํ‰๊ฐ€์˜ ํ•„์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค.
โ€ข
๋ฒค์น˜๋งˆํฌ ๊ตฌ์ถ• ๊ณผ์ •์˜ ๋ณต์žก์„ฑ๊ณผ ๋น„์šฉ, ๊ทธ๋ฆฌ๊ณ  ๋‹ค์–‘ํ•œ ๋ฌธํ™”๊ถŒ์˜ ๋ฏธ๋ฌ˜ํ•œ ์ฐจ์ด๋ฅผ ์™„์ „ํžˆ ํฌ์ฐฉํ•˜๋Š” ๋ฐ์˜ ์–ด๋ ค์›€์ด ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.
โ€ข
ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ๋” ๋งŽ์€ ์–ธ์–ด์™€ ๋ฌธํ™”๊ถŒ์œผ๋กœ ๋ฒค์น˜๋งˆํฌ๋ฅผ ํ™•์žฅํ•˜๊ณ , LLM์ด ๋ฌธํ™”์  ๋งฅ๋ฝ์„ ์ดํ•ดํ•˜๋Š” ์‹ฌ์ธต์ ์ธ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํƒ๊ตฌํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
๐Ÿ‘