FunctionalAgent: Towards end-to-end on-top functional design

์ €์ž: Yuhao Chen, Donald G. Truhlar, Xiao He | ๋‚ ์งœ: 2026-05-07 | URL: https://arxiv.org/abs/2605.06215 📄 PDF


Essence

FunctionalAgent๋Š” LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์œผ๋กœ MC-PDFT์˜ on-top functional ์ž๋™ ๊ฐœ๋ฐœ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•œ๋‹ค. ๋ฐ์ดํ„ฐ์…‹ ๊ตฌ์ถ•, active-space ์ƒ์„ฑ, MCSCF ๊ณ„์‚ฐ, descriptor ์ƒ์„ฑ, loss-function ๊ตฌ์„ฑ, functional fitting ๋ฐ ํ‰๊ฐ€๋ฅผ ํ†ตํ•ฉํ•˜๋Š” end-to-end ์ž๋™ํ™” ์›Œํฌํ”Œ๋กœ์šฐ๋ฅผ ์ œ์‹œํ•œ๋‹ค.

Motivation

Achievement

Figure 2

Figure 2. LLM-assisted workflow for constructing and refining quantum-chemical benchmark datasets. a,

How

Figure 2

Figure 2. LLM-assisted workflow for constructing and refining quantum-chemical benchmark datasets. a,

Originality

Limitation & Further Study

ํ›„์†์—ฐ๊ตฌ: (1) Larger scale์˜ ํ™”ํ•™ ์—๋„ˆ์ง€ ๋ฐ์ดํ„ฐ์…‹์—์„œ์˜ ์„ฑ๋Šฅ ๊ฒ€์ฆ; (2) Functional generalization๊ณผ transferability์˜ ๋” ๊นŠ์€ ๋ถ„์„; (3) LLM agent์˜ decision-making process ํ•ด์„์„ฑ ํ–ฅ์ƒ; (4) ๋‹ค๋ฅธ quantum chemistry methods (์˜ˆ: Post-Hartree-Fock, DFT)์—์˜ ์ ์šฉ ํ™•๋Œ€

Evaluation

Novelty: 4/5 Technical Soundness: 4/5 Significance: 4/5 Clarity: 4/5 Overall: 4/5

์ดํ‰: ๋ณธ ๋…ผ๋ฌธ์€ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์„ ์–‘์žํ™”ํ•™์˜ functional ๊ฐœ๋ฐœ ์ž๋™ํ™”์— ํ˜์‹ ์ ์œผ๋กœ ์ ์šฉํ•˜์—ฌ, end-to-end closed-loop workflow๋ฅผ ํ†ตํ•ด MC26๊ณผ COF26์ด๋ผ๋Š” ๊ฐœ์„ ๋œ on-top functionals์„ ์„ฑ๊ณต์ ์œผ๋กœ ๊ฐœ๋ฐœํ–ˆ๋‹ค. ๊ณ„์ธต์  ์•„ํ‚คํ…์ฒ˜์™€ constrained agent design์œผ๋กœ controllability๋ฅผ ํ™•๋ณดํ•˜๋ฉด์„œ๋„ ๋ณต์žกํ•œ scientific workflow๋ฅผ ์ž๋™ํ™”ํ•œ ์ ์ด ๋†’์ด ํ‰๊ฐ€๋˜๋‚˜, ์ •๋Ÿ‰์  ํšจ์œจ ์„ฑ๊ณผ์™€ ๊ด‘๋ฒ”์œ„ํ•œ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ๋” ์ƒ์„ธํ•œ ๊ฒ€์ฆ์ด ์š”๊ตฌ๋œ๋‹ค.

๊ฐ™์ด ๋ณด๋ฉด ์ข‹์€ ๋…ผ๋ฌธ

๊ธฐ๋ฐ˜ ์—ฐ๊ตฌ
308๋ฒˆ ๋…ผ๋ฌธ์€ ์–‘์žํ™”ํ•™ ์‹คํ—˜ ์ž๋™ํ™”๋ฅผ ์œ„ํ•œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋‹ค๋ฃจ์–ด, 3110์˜ on-top ํ•จ์ˆ˜ ์„ค๊ณ„ ์ž๋™ํ™”์˜ ๊ทผ๊ฑฐ ๋ฐ ์‹คํ˜„ ๊ฐ€๋Šฅ์„ฑ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๋‹ค๋ฅธ ์ ‘๊ทผ
PhyNiKCE๋Š” ๋‰ด๋กœ์‹ฌ๋ณผ๋ฆญ AI๋ฅผ ํ†ตํ•œ ๊ณผํ•™์  ํ•จ์ˆ˜ ์„ค๊ณ„ ์—์ด์ „ํŠธ ํ”„๋ ˆ์ž„์›Œํฌ๋กœ, 3110์˜ LLM ๊ธฐ๋ฐ˜ ์ž๋™ ์„ค๊ณ„์™€ ๋ฌธ์ œ ์„ค์ •์ด ์œ ์‚ฌํ•˜๋‹ค.
๋‹ค๋ฅธ ์ ‘๊ทผ
Density Functional์˜ ์ž๋™ ์„ค๊ณ„ ์ „๋žต์„ LLM ๊ธฐ๋ฐ˜ ์‹œ์Šคํ…œ์œผ๋กœ ๊ตฌํ˜„ํ•˜๋Š” ๋น„์Šทํ•œ but slightly different agentic ์ ‘๊ทผ ์‚ฌ๋ก€์ž…๋‹ˆ๋‹ค.
ํ›„์† ์—ฐ๊ตฌ
2199๋ฒˆ ๋…ผ๋ฌธ์€ ๊ณผํ•™์  AI agent์˜ ์ด๋ก ๊ณผ ์‹ค์ œ ๊ตฌํ˜„ ์‚ฌ๋ก€๋ฅผ ๊ฐœ๊ด„ํ•˜์—ฌ, 3110์˜ end-to-end ์ž๋™ํ™” ์›Œํฌํ”Œ๋กœ์šฐ์˜ ์ „ํ†ต์  ๊ฐ€์น˜์™€ ํ–ฅํ›„ ๋ฐœ์ „ ๋ฐฉํ–ฅ์„ ๋…ผ์˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์‘์šฉ ์‚ฌ๋ก€
๋จธ์‹ ๋Ÿฌ๋‹ ์ƒํ˜ธ์ž‘์šฉ ํผํ…์…œ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ ํšจ์œจ์  ๋ฏธ์„ธ์กฐ์ • ์—ฐ๊ตฌ๋กœ, FunctionalAgent์˜ ์›Œํฌํ”Œ๋กœ์šฐ๊ฐ€ ML ํผํ…์…œ ์‹ ๋ขฐ์„ฑ ๊ฐœ์„ ์— ์‹ค์ œ ์ ์šฉ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
← ๋ชฉ๋ก์œผ๋กœ ๋Œ์•„๊ฐ€๊ธฐ

๐ŸŽง Audio Overview

์ด ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ๋ฅผ ํŒŸ์บ์ŠคํŠธํ˜• ์˜ค๋””์˜ค๋กœ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. (Gemini ยท ํ‚ค๋Š” ๋ธŒ๋ผ์šฐ์ €์—๋งŒ ์ €์žฅ ยท ์™„์„ฑ๋ณธ์€ ์ด๋ฉ”์ผ๋กœ๋„ ์ „์†ก)
โ–ธ ๊ณ ๊ธ‰: ๊ตฌ์„ฑ ๋ฐฉํ–ฅ(๋Œ€๋ณธ ์ž‘์„ฑ ์ง€์นจ) ์ง์ ‘ ์ˆ˜์ •