Sign In

PATHWAYS: Evaluating Investigation and Context Discovery in AI Web Agents

Created by
  • Haebom
Category
Empty

์ €์ž

Shifat E. Arman, Syed Nazmus Sakib, Tapodhir Karmakar Taton, Nafiul Haque, Shahrear Bin Amin

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ ์›น ๊ธฐ๋ฐ˜ AI ์—์ด์ „ํŠธ๊ฐ€ ์ˆจ๊ฒจ์ง„ ๋งฅ๋ฝ ์ •๋ณด๋ฅผ ๋ฐœ๊ฒฌํ•˜๊ณ  ํ™œ์šฉํ•˜๋Š” ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ 250๊ฐ€์ง€ ๋‹ค๋‹จ๊ณ„ ์˜์‚ฌ ๊ฒฐ์ • ์ž‘์—…์œผ๋กœ ๊ตฌ์„ฑ๋œ PATHWAYS ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ ๊ฒฐ๊ณผ, ์—์ด์ „ํŠธ๋“ค์€ ๊ด€๋ จ ํŽ˜์ด์ง€๋ฅผ ์ž˜ ํƒ์ƒ‰ํ•˜์ง€๋งŒ, ๊ฒฐ์ •์ ์ธ ์ˆจ๊ฒจ์ง„ ์ฆ๊ฑฐ๋ฅผ ์ฐพ๋Š” ๋น„์œจ์€ ๋‚ฎ์•˜์œผ๋ฉฐ, ์˜ค๋„ํ•˜๋Š” ํ‘œ๋ฉด์  ์‹ ํ˜ธ๋ฅผ ๊ทน๋ณตํ•ด์•ผ ํ•˜๋Š” ์ž‘์—…์—์„œ๋Š” ์ •ํ™•๋„๊ฐ€ ํ˜„์ €ํžˆ ๋–จ์–ด์กŒ์Šต๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
์›น ์—์ด์ „ํŠธ๊ฐ€ ํƒ์ƒ‰ ๊ณผ์ •์—์„œ ๋งฅ๋ฝ ์ •๋ณด๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ๋ฐœ๊ฒฌํ•˜๊ณ  ํ™œ์šฉํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
โ€ข
์—์ด์ „ํŠธ๋“ค์ด ์ž˜๋ชป๋œ ์ •๋ณด์— ๊ธฐ๋ฐ˜ํ•œ ์ถ”๋ก ์„ ์ƒ์„ฑํ•˜๊ฑฐ๋‚˜, ๋ฐœ๊ฒฌํ•œ ์ •๋ณด๋ฅผ ์ตœ์ข… ์˜์‚ฌ ๊ฒฐ์ •์— ํ†ตํ•ฉํ•˜์ง€ ๋ชปํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ์ง€์ ํ•ฉ๋‹ˆ๋‹ค.
โ€ข
๋ช…์‹œ์ ์ธ ์ง€์‹œ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฒƒ์ด ๋งฅ๋ฝ ๋ฐœ๊ฒฌ์—๋Š” ๋„์›€์ด ๋˜์ง€๋งŒ, ์ „๋ฐ˜์ ์ธ ์ •ํ™•๋„๋ฅผ ์ €ํ•ดํ•  ์ˆ˜ ์žˆ์–ด ์ ˆ์ฐจ ์ค€์ˆ˜์™€ ํšจ๊ณผ์ ์ธ ํŒ๋‹จ ์‚ฌ์ด์˜ ์ ˆ์ถฉ์ ์ด ์กด์žฌํ•จ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.
โ€ข
ํ˜„์žฌ ์›น ์—์ด์ „ํŠธ ๊ตฌ์กฐ๋Š” ์ ์‘์  ํƒ์ƒ‰, ์ฆ๊ฑฐ ํ†ตํ•ฉ, ํŒ๋‹จ ์žฌ๊ณ ๋ฅผ ์œ„ํ•œ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ๋ถ€์กฑํ•˜๋ฉฐ, ์ด๋Š” ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ ์ด๋Ÿฌํ•œ ๋ถ€๋ถ„์„ ๊ฐœ์„ ํ•  ํ•„์š”๊ฐ€ ์žˆ์Œ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
๐Ÿ‘