Sign In

Many-Shot CoT-ICL: Making In-Context Learning Truly Learn

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

μ €μž

Tsz Ting Chung, Lemao Liu, Mo Yu, Dit-Yan Yeung

πŸ’‘ κ°œμš”

λ³Έ μ—°κ΅¬λŠ” μΆ”λ‘  μž‘μ—…μ—μ„œ λ‹€μˆ˜μ˜ μ˜ˆμ‹œλ₯Ό ν™œμš©ν•˜λŠ” Chain-of-Thought In-Context Learning(CoT-ICL)의 μž‘λ™ 방식을 νƒκ΅¬ν•©λ‹ˆλ‹€. 연ꡬ진은 CoT-ICL을 λ‹¨μˆœ νŒ¨ν„΄ 맀칭이 μ•„λ‹Œ 'In-Context Test-Time Learning'으둜 ν•΄μ„ν•˜κ³ , μ‰¬μš΄ 이해와 κ°œλ…μ  점진적 λ°œμ „μ„ μ§€μ›ν•˜λŠ” μ˜ˆμ‹œ 선택 및 μˆœμ„œ κ²°μ •μ˜ μ€‘μš”μ„±μ„ κ°•μ‘°ν•©λ‹ˆλ‹€. 이λ₯Ό 기반으둜 μ œμ•ˆλœ Curvilinear Demonstration Selection (CDS) 방법은 μˆ˜ν•™ κ³Όμ œμ—μ„œ 64개의 μ˜ˆμ‹œ μ‚¬μš© μ‹œ μ΅œλŒ€ 5.42%p의 μ„±λŠ₯ ν–₯상을 λ³΄μ—¬μ£Όμ—ˆμŠ΅λ‹ˆλ‹€.

πŸ”‘ μ‹œμ‚¬μ  및 ν•œκ³„

β€’
λ§Žμ€ μ˜ˆμ‹œλ₯Ό μ‚¬μš©ν•˜λŠ” In-Context Learning (ICL)이 λ‹¨μˆœνžˆ 정보λ₯Ό κ²€μƒ‰ν•˜λŠ” 것이 μ•„λ‹ˆλΌ, λͺ¨λΈμ΄ ν•™μŠ΅ν•˜λŠ” κ³Όμ •μž„μ„ μ‹œμ‚¬ν•©λ‹ˆλ‹€.
β€’
μΆ”λ‘  μž‘μ—…μ—μ„œ CoT-ICL의 효과λ₯Ό κ·ΉλŒ€ν™”ν•˜κΈ° μœ„ν•΄μ„œλŠ” μ˜ˆμ‹œμ˜ λ‚΄μš©κ³Ό μˆœμ„œκ°€ 맀우 μ€‘μš”ν•˜λ©°, 체계적인 '컀리큘럼'처럼 κ΅¬μ„±λ˜μ–΄μ•Ό ν•©λ‹ˆλ‹€.
β€’
μ œμ•ˆλœ Curvilinear Demonstration Selection (CDS) 방법은 μ˜ˆμ‹œμ˜ μˆœμ„œλ₯Ό μ‘°μ ˆν•˜μ—¬ μ„±λŠ₯을 ν–₯μƒμ‹œν‚¬ 수 μžˆλŠ” 효과적인 λ°©λ²•μž„μ„ λ³΄μ—¬μ€λ‹ˆλ‹€.
β€’
λ‹€μ–‘ν•œ μ’…λ₯˜μ˜ μΆ”λ‘  μž‘μ—…κ³Ό λͺ¨λΈμ— λŒ€ν•œ CDS λ°©λ²•μ˜ μΌλ°˜ν™” κ°€λŠ₯성을 μΆ”κ°€μ μœΌλ‘œ 검증할 ν•„μš”κ°€ μžˆμŠ΅λ‹ˆλ‹€.
πŸ‘
Made with Slashpage