Sign In

The Synthetic Web: Adversarially-Curated Mini-Internets for Diagnosing Epistemic Weaknesses of Language Agents

Created by
  • Haebom
Category
Empty

์ €์ž

Shrey Shah, Levent Ozgur

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ ์›น ๊ฒ€์ƒ‰ ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ˜ ์–ธ์–ด ์—์ด์ „ํŠธ๋“ค์ด ์‹ ๋ขฐํ•  ์ˆ˜ ์—†๊ฑฐ๋‚˜ ์ ๋Œ€์ ์ธ ์ •๋ณด์— ์ทจ์•ฝํ•˜๋‹ค๋Š” ๋ฌธ์ œ์ ์„ ์ง€์ ํ•˜๋ฉฐ, ์ด๋ฅผ ์ง„๋‹จํ•˜๊ธฐ ์œ„ํ•œ 'ํ•ฉ์„ฑ ์›น ๋ฒค์น˜๋งˆํฌ'๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ๋Š” ์ ˆ์ฐจ์ ์œผ๋กœ ์ƒ์„ฑ๋œ ๊ฐ€์ƒ์˜ ์›น ํ™˜๊ฒฝ์—์„œ ์˜๋„์ ์œผ๋กœ ์กฐ์ž‘๋œ ๊ฒ€์ƒ‰ ์ˆœ์œ„๋ฅผ ํ†ตํ•ด ์–ธ์–ด ์—์ด์ „ํŠธ์˜ ์ทจ์•ฝ์„ฑ์„ ์ฒด๊ณ„์ ์œผ๋กœ ์ธก์ •ํ•˜๋ฉฐ, ์‹คํ—˜ ๊ฒฐ๊ณผ ์ตœ์‹  ์–ธ์–ด ๋ชจ๋ธ๋“ค์ด ์ž˜๋ชป๋œ ์ •๋ณด์— ์‰ฝ๊ฒŒ ํ˜„ํ˜น๋˜์–ด ์‹ฌ๊ฐํ•œ ์„ฑ๋Šฅ ์ €ํ•˜๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
์ตœ์‹  ์–ธ์–ด ๋ชจ๋ธ๋“ค์€ ์‹ ๋ขฐํ•  ์ˆ˜ ์—†๋Š” ์ •๋ณด์— ๋Œ€ํ•œ ์ธ์ง€์  ์ทจ์•ฝ์„ฑ์ด ๋งค์šฐ ๋†’์•„, ๊ณ ์œ„ํ—˜ ํ™˜๊ฒฝ์—์„œ์˜ ๋ฐฐํฌ์— ๋Œ€ํ•œ ๊ทผ๋ณธ์ ์ธ ์žฌ๊ณ ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
โ€ข
์ œ์•ˆ๋œ 'ํ•ฉ์„ฑ ์›น ๋ฒค์น˜๋งˆํฌ'๋Š” ์–ธ์–ด ์—์ด์ „ํŠธ์˜ ๊ฒ€์ƒ‰ ์ ๋Œ€์„ฑ์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ๋ถ„์„๊ณผ ๋ฐฉ์–ด ์ „๋žต ํ‰๊ฐ€๋ฅผ ์œ„ํ•œ ์ค‘์š”ํ•œ ๋„๊ตฌ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
โ€ข
ํ–ฅํ›„ ์—ฐ๊ตฌ๋Š” ์ด๋Ÿฌํ•œ ์ทจ์•ฝ์ ์„ ๊ทน๋ณตํ•˜๊ณ  ์กฐ์ž‘์— ์ €ํ•ญํ•˜๋ฉฐ, ๋ถˆํ™•์‹ค์„ฑ์„ ์ธ์ง€ํ•˜๋Š”(epistemically humble) ์—์ด์ „ํŠธ ๊ฐœ๋ฐœ์— ์ดˆ์ ์„ ๋งž์ถฐ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๐Ÿ‘