2026-05-25

๋ฐฐ๊ฒฝ์€ ๋น ๋ฅธ ์ƒํƒœ๊ณ„ ๋ณ€ํ™”์˜ˆ์š”

๐Ÿ’ก ํ•œ์ค„ ์š”์•ฝ|๋ฐฐ๊ฒฝ์€ ๋น ๋ฅธ ์ƒํƒœ๊ณ„ ๋ณ€ํ™”์˜ˆ์š”.

ํ—ˆ๊น…ํŽ˜์ด์Šค๊ฐ€ AI ์—์ด์ „ํŠธ ์šฉ์–ด ํ˜ผ์„ ์„ ์ค„์ด๊ธฐ ์œ„ํ•œ ๊ณต์‹ ์šฉ์–ด์ง‘์„ ๊ณต๊ฐœํ–ˆ์–ด์š”.

๋ฐฐ๊ฒฝ์€ ๋น ๋ฅธ ์ƒํƒœ๊ณ„ ๋ณ€ํ™”์˜ˆ์š”. ICLR 2026 ์ดํ›„ ํ˜„์—…์—์„œ๋„ harness์™€ scaffold ๊ฐ™์€ ๋‹จ์–ด ์˜๋ฏธ๊ฐ€ ์ œ๊ฐ๊ฐ์ด๋ผ๋Š” ๋ฌธ์ œ ์ œ๊ธฐ๊ฐ€ ๋‚˜์™”๊ณ , ๊ฐ™์€ ์šฉ์–ด๋ฅผ ํ”„๋ ˆ์ž„์›Œํฌ๋งˆ๋‹ค ๋‹ค๋ฅด๊ฒŒ ์“ฐ๋Š” ์ƒํ™ฉ์ด ์ด์–ด์กŒ๊ฑฐ๋“ ์š”.

์ด๋ฒˆ ๊ธ€์€ Model, Scaffolding, Harness, Agent, Context Engineering, Policy, Tool Use, Skills, Sub-agents๊นŒ์ง€ 9๊ฐœ ํ•ต์‹ฌ ๊ฐœ๋…๊ณผ, ํ•™์Šต ํŒŒํŠธ์˜ RL EnvironmentยทTrainerยทRolloutยทReward 4๊ฐœ๊นŒ์ง€ ์ •๋ฆฌํ–ˆ์–ด์š”. ์šฉ์–ด ํ‘œ์ค€์„ ๊ฐ•์ œํ•˜๊ธฐ๋ณด๋‹ค ์‹ค๋ฌด ๊ณตํ†ต ์ดํ•ด๋ฅผ ๋งž์ถ”๋Š” ๋ฐ ์ดˆ์ ์„ ๋‘” ์ ์ด ํฌ์ธํŠธ์˜ˆ์š”.

์ด ๊ธ€์ด ์–ด๋• ๋‚˜์š”?

๊ด€๋ จ ๊ธ€

๊ณต๊ฐœ ๋ณด์•ˆ ์Šค์บ๋„ˆ๋Š” ์‹ค์ œ ๋ฐฉ๋ฌธ์ž๊ฐ€ ๋ณผ ์ˆ˜ ์žˆ๋Š” ๊ฒƒ๋งŒ ๊ฒ€์‚ฌํ•ด์š”

๊ณต๊ฐœ ๋ณด์•ˆ ์Šค์บ๋„ˆ๋Š” ์‹ค์ œ ๋ฐฉ๋ฌธ์ž๊ฐ€ ๋ณผ ์ˆ˜ ์žˆ๋Š” ๊ฒƒ๋งŒ ๊ฒ€์‚ฌํ•ด์š”.

์žก๋Œ์Œค2๋ถ„ ์†Œ์š”

์ค‘๊ตญ ์‚ฌ์ด๋ฒ„๊ณต๊ฐ„๊ด€๋ฆฌ๊ตญ(CAC)์ด ์•Œ๋ฆฌ๋ฐ”๋ฐ”์˜ Qwen AI ๋ชจ๋ธ์„ ํƒ‘์žฌํ•œ ์• ํ”Œ์˜ AI ์„œ๋น„์Šค๋ฅผ ๊ณต์‹ ์Šน์ธํ–ˆ๊ฑฐ๋“ ์š”

์ค‘๊ตญ ์‚ฌ์ด๋ฒ„๊ณต๊ฐ„๊ด€๋ฆฌ๊ตญ(CAC)์ด ์•Œ๋ฆฌ๋ฐ”๋ฐ”์˜ Qwen AI ๋ชจ๋ธ์„ ํƒ‘์žฌํ•œ ์• ํ”Œ์˜ AI ์„œ๋น„์Šค๋ฅผ ๊ณต์‹ ์Šน์ธํ–ˆ๊ฑฐ๋“ ์š”.

์žก๋Œ์Œค2๋ถ„ ์†Œ์š”

์ œ๋ฏธ๋‚˜์ด 2.5 ํ”„๋กœ ์ดํ›„ ํ›„์† ํ”Œ๋ž˜๊ทธ์‹ญ ์—†์ด ์ง€๋‚ด์˜จ ๊ตฌ๊ธ€์ด ์ด๋ฒˆ์—” ์ œ๋ฏธ๋‚˜์ด 3.5 ํ”„๋กœ๋กœ ๋ฐ˜๊ฒฉ์„ ์˜ˆ๊ณ ํ•˜๊ฑฐ๋“ ์š”

์ œ๋ฏธ๋‚˜์ด 2.5 ํ”„๋กœ ์ดํ›„ ํ›„์† ํ”Œ๋ž˜๊ทธ์‹ญ ์—†์ด ์ง€๋‚ด์˜จ ๊ตฌ๊ธ€์ด ์ด๋ฒˆ์—” ์ œ๋ฏธ๋‚˜์ด 3.5 ํ”„๋กœ๋กœ ๋ฐ˜๊ฒฉ์„ ์˜ˆ๊ณ ํ•˜๊ฑฐ๋“ ์š”.

์žก๋Œ์Œค2๋ถ„ ์†Œ์š”