Sign In

Do Multi-Agents Dream of Electric Screens? Achieving Perfect Accuracy on AndroidWorld Through Task Decomposition

Created by
  • Haebom
Category
Empty

์ €์ž

Pierre-Louis Favreau, Jean-Pierre Lo, Clement Guiguet, Charles Simon-Meunier, Nicolas Dehandschoewercker, Allen G. Roush, Judah Goldfeder, Ravid Shwartz-Ziv

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ AndroidWorld ๋ฒค์น˜๋งˆํฌ์˜ 116๊ฐœ ๋ชจ๋“  ๊ณผ์ œ๋ฅผ 100% ์„ฑ๊ณต์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ 'Minitap'์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด ๋‹จ์ผ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์˜ ์‹คํŒจ ์›์ธ(๋งฅ๋ฝ ์˜ค์—ผ, ํ…์ŠคํŠธ ์ž…๋ ฅ ์‹คํŒจ, ๋ฐ˜๋ณต ๋ฃจํ”„)์„ ๋ถ„์„ํ•˜๊ณ , ์ธ์ง€์  ๋ถ„๋ฆฌ, ๊ฒฐ์ •๋ก ์  ํ…์ŠคํŠธ ์ž…๋ ฅ ๊ฒ€์ฆ, ๋ฉ”ํƒ€์ธ์ง€์  ์ถ”๋ก ์„ ํ†ตํ•ด ์ด๋ฅผ ํ•ด๊ฒฐํ–ˆ์Šต๋‹ˆ๋‹ค. Minitap์€ 80%์˜ ์ธ๊ฐ„ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ, ์˜คํ”ˆ ์†Œ์Šค๋กœ ๊ณต๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
๋ณต์žกํ•œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์ด ๋‹จ์ผ ์—์ด์ „ํŠธ๋ณด๋‹ค ํ›จ์”ฌ ํšจ๊ณผ์ ์ด๋ฉฐ, ๊ฐ ์—์ด์ „ํŠธ์˜ ์ „๋ฌธํ™”๋ฅผ ํ†ตํ•ด ์„ฑ๋Šฅ์„ ๊ทน๋Œ€ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
โ€ข
ํ…์ŠคํŠธ ์ž…๋ ฅ์˜ ์ •ํ™•์„ฑ์„ ๊ธฐ๊ธฐ ์ƒํƒœ์™€ ๋น„๊ตํ•˜์—ฌ ๊ฒ€์ฆํ•˜๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ์—์ด์ „ํŠธ์˜ ์‹ ๋ขฐ์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
โ€ข
ํ˜„์žฌ์˜ ์ ‘๊ทผ ๋ฐฉ์‹์ด AndroidWorld ๋ฒค์น˜๋งˆํฌ์— ํŠนํ™”๋˜์–ด ์žˆ์–ด, ๋‹ค๋ฅธ ๋ณต์žกํ•œ ์‹ค์ œ ํ™˜๊ฒฝ์ด๋‚˜ ๋‹ค๋ฅธ ์ข…๋ฅ˜์˜ ํƒœ์Šคํฌ๋กœ์˜ ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ์„ฑ์€ ์ถ”๊ฐ€์ ์ธ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
๐Ÿ‘