Sign In

Multi-Paradigm Agent Interaction in Practice:A Systematic Analysis of Generator-Evaluator, ReAct Loop,and Adversarial Evaluation in the buddyMe Framework

μž‘μ„±μž
  • Haebom
μΉ΄ν…Œκ³ λ¦¬
Empty

μ €μž

Xiaohua Wang, Chao Han, Kai Yu, XiaoLiang Xu, Liang Wang

πŸ’‘ κ°œμš”

λ³Έ 논문은 LLM μ—μ΄μ „νŠΈμ˜ λ‹€μ–‘ν•œ μƒν˜Έμž‘μš© νŒ¨λŸ¬λ‹€μž„ 쀑 Generator-Evaluator, ReAct Loop, Memory-Augmented Interaction을 buddyMe ν”„λ ˆμž„μ›Œν¬μ— ν†΅ν•©ν•˜κ³  μ²΄κ³„μ μœΌλ‘œ λΆ„μ„ν•©λ‹ˆλ‹€. 5단계 처리 νŒŒμ΄ν”„λΌμΈκ³Ό 6차원 평가 체계λ₯Ό 톡해, λ‹€μ–‘ν•œ μ‹€μ œ 적용 사둀λ₯Ό λΆ„μ„ν•œ κ²°κ³Ό Generator-Evaluator 사전 κ²€ν† κ°€ μš”κ΅¬μ‚¬ν•­ λˆ„λ½μ„ 효과적으둜 λ°œκ²¬ν•˜κ³ , ReAct λ£¨ν”„λŠ” μ•ˆμ •μ μΈ ν•˜μœ„ νƒœμŠ€ν¬ 싀행을 보μž₯ν•˜μ§€λ§Œ 도ꡬ 호좜 쀑볡이 λ°œμƒν•˜λ©°, Adversarial Evaluation은 주둜 λ‚΄μš© μ •μ œμ— 기여함을 λ°ν˜”μŠ΅λ‹ˆλ‹€.

πŸ”‘ μ‹œμ‚¬μ  및 ν•œκ³„

β€’
Generator-Evaluator 방식은 λ³΅μž‘ν•œ νƒœμŠ€ν¬μ—μ„œ μš”κ΅¬μ‚¬ν•­ λˆ„λ½μ„ 20% μˆ˜μ€€μœΌλ‘œ νƒμ§€ν•˜μ—¬ 초기 μ κ²€μ˜ νš¨μœ¨μ„±μ„ λ†’μž…λ‹ˆλ‹€.
β€’
ReAct λ£¨ν”„λŠ” ν•˜μœ„ νƒœμŠ€ν¬ μ‹€ν–‰μ˜ μ•ˆμ •μ„±μ„ ν™•λ³΄ν•˜μ§€λ§Œ, μ•½ 30%의 λΆˆν•„μš”ν•œ 도ꡬ ν˜ΈμΆœμ„ μœ λ°œν•˜λŠ” 단점이 μžˆμŠ΅λ‹ˆλ‹€.
β€’
Adversarial Evaluation λ©”μ»€λ‹ˆμ¦˜μ€ 2-3번의 λŒ€ν™”λ‘œ μ•½ 70%의 μ‹œλ‚˜λ¦¬μ˜€μ—μ„œ ν•©μ˜μ— λ„λ‹¬ν•˜λ©°, 논리적 였λ₯˜ μˆ˜μ •λ³΄λ‹€λŠ” λ‚΄μš© μ •μ œμ— 더 νš¨κ³Όμ μž…λ‹ˆλ‹€.
β€’
ν˜„μž¬ μ—°κ΅¬λŠ”buddyMe ν”„λ ˆμž„μ›Œν¬ λ‚΄μ—μ„œμ˜ 뢄석에 μ΄ˆμ μ„ λ§žμΆ”κ³  있으며, λ‹€μ–‘ν•œ ν”„λ ˆμž„μ›Œν¬μ™€μ˜ 비ꡐλ₯Ό 톡해 μΌλ°˜ν™”λœ 닀쀑 νŒ¨λŸ¬λ‹€μž„ μ—μ΄μ „νŠΈ μ‹œμŠ€ν…œ 섀계 지침을 μ œκ³΅ν•˜μ§€λ§Œ, μ‹€μ œ λ³΅μž‘ν•œ ν™˜κ²½μ—μ„œμ˜ ν™•μž₯μ„± 및 λ‹€μ–‘ν•œ μ’…λ₯˜μ˜ 였λ₯˜μ— λŒ€ν•œ 좔가적인 연ꡬ가 ν•„μš”ν•©λ‹ˆλ‹€.
πŸ‘
Made with Slashpage