yunho

yunho

์ฑ—๋ด‡์„ ๋ฐ‘๋ฐ”๋‹ฅ๋ถ€ํ„ฐ ๋งŒ๋“œ๋Š” ๊ธฐ๋ก

๐Ÿ“„ ๋…ผ๋ฌธ๋ฆฌ๋ทฐ

์ฑ—๋ด‡ยทNLPยทLLMยทAgent ๊ด€๋ จ ๋…ผ๋ฌธ์„ ์ฝ๊ณ  ํ•ต์‹ฌ๋งŒ ์ •๋ฆฌํ•˜๋Š” ๊ณต๊ฐ„์ž…๋‹ˆ๋‹ค.

Voyager Paper review

๋งˆ์ธํฌ๋ž˜ํ”„ํŠธ์—์„œ ์Šค์Šค๋กœ ๊ณผ์ œ๋ฅผ ์ •ํ•˜๊ณ , ์„ฑ๊ณตํ•œ ํ–‰๋™์„ ์ฝ”๋“œ๋กœ ์ €์žฅํ•ด ์žฌ์‚ฌ์šฉํ•˜๋Š” lifelong learning ์—์ด์ „ํŠธ. ReAct์™€ Reflexion์€ ์—ฌ๊ธฐ์„œ ๋‚˜๋ฌด ๋„๊ตฌ๋„ ๋ชป ๋งŒ๋“ค์—ˆ๋‹ค.

LLM Agent Survey Paper review

ProfileยทMemoryยทPlanningยทAction, ์—์ด์ „ํŠธ 4๋ชจ๋“ˆ์ด๋ผ๋Š” ํ†ต๋…์˜ ์ถœ์ฒ˜. CoALA๊ฐ€ ์ด๋ก ์ด๋ผ๋ฉด ์ด ์„œ๋ฒ ์ด๋Š” 2023๋…„๊นŒ์ง€์˜ ์—์ด์ „ํŠธ ์—ฐ๊ตฌ ์ „์ฒด๋ฅผ ์ •๋ฆฌํ•œ ์นดํƒˆ๋กœ๊ทธ๋‹ค.

CoALA Paper review

์—์ด์ „ํŠธ๋ฅผ ๋ณผ ๋•Œ ์„ธ ๊ฐ€์ง€๋งŒ ๋ฌผ์œผ๋ฉด ๋œ๋‹ค. ๋ฌด์—‡์„ ๊ธฐ์–ตํ•˜๋Š”๊ฐ€, ๋ฌด์—‡์„ ํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€, ์–ด๋–ป๊ฒŒ ๊ฒฐ์ •ํ•˜๋Š”๊ฐ€. ReAct๋ถ€ํ„ฐ Voyager๊นŒ์ง€๋ฅผ ์ด ์„ธ ์งˆ๋ฌธ์œผ๋กœ ์ •๋ฆฌํ•œ ๋…ผ๋ฌธ.

Generative Agents Paper review

LLM ์—์ด์ „ํŠธ 25๋ช…์„ ๊ฐ€์ƒ ๋งˆ์„์— ํ’€์–ด๋†“์œผ๋ฉด ๋ฌด์Šจ ์ผ์ด ๋ฒŒ์–ด์ง€๋Š”๊ฐ€. ํ–‰๋™์˜ ์ •ํ™•๋„๊ฐ€ ์•„๋‹ˆ๋ผ, ๊ธฐ์–ต-ํšŒ์ƒ-๋ฐ˜์„ฑ-๊ณ„ํš ๊ตฌ์กฐ๊ฐ€ ๊ฒฝํ—˜์— ๊ทผ๊ฑฐํ•œ ๋ฏฟ์„ ๋งŒํ•œ ํ–‰๋™์„ ์œ ์ง€์‹œํ‚ค๋Š”์ง€๋ฅผ ์ธก์ •ํ•œ ๋…ผ๋ฌธ.

Toolformer Paper review

์‚ฌ๋žŒ ์–ด๋…ธํ…Œ์ด์…˜ ์—†์ด LM์ด ์Šค์Šค๋กœ ๋„๊ตฌ ์‚ฌ์šฉ๋ฒ•์„ ๋ฐฐ์šด๋‹ค. โ€˜์ด ๋„๊ตฌ ํ˜ธ์ถœ์ด ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก์„ ์‰ฝ๊ฒŒ ํ•˜๋Š”๊ฐ€โ€™๋ผ๋Š” ํ•„ํ„ฐ ํ•˜๋‚˜๋กœ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ๋งŒ๋“  ๋…ผ๋ฌธ.

Reflexion Paper review

์‹คํŒจ๋ฅผ ๋ฐ˜์„ฑ๋ฌธ์œผ๋กœ ๋‚จ๊ฒจ ๋‹ค์Œ ์‹œ๋„์— ์จ๋จน๋Š” ์—์ด์ „ํŠธ. ๊ฐ€์ค‘์น˜ ์—…๋ฐ์ดํŠธ ์—†๋Š” โ€˜๋ง๋กœ ํ•˜๋Š” ๊ฐ•ํ™”ํ•™์Šตโ€™์œผ๋กœ HumanEval 91%๋ฅผ ์ฐ์—ˆ๋‹ค.

ReAct Paper review

์ง€๊ธˆ ๋ชจ๋“  LLM Agent์˜ ๋ผˆ๋Œ€๊ฐ€ ๋œ ๋…ผ๋ฌธ. Thought โ†’ Action โ†’ Observation ๋ฃจํ”„๋ถ€ํ„ฐ LangChain create_agent๊นŒ์ง€ ๋œฏ์–ด๋ณธ๋‹ค.

SOLAR Paper review

SOLAR 10.7B: Scaling Large Language Models with Simple yet Effective Depth Up-Scaling Paper review

PPTOD paper review

Multi-Task Pre-Training for Plug-and-Play Task-Oriented Dialogue System

UBAR paper review

UBAR: Towards Fully End-to-End Task-Oriented Dialog System with GPT-2

DSS-DST Paper review

Dual Slot Selector via Local Reliability Verification for Dialogue State Tracking

TripPy + SaCLog Paper review

Preview, Attend and Review: Schema-Aware Curriculum Learning for Multi-Domain Dialog State Tracking

TripPy Paper review

TripPy: A Triple Copy Strategy for Value Independent Neural Dialog State Tracking

SOLOIST paper review

SOLOIST: Building Task Bots at Scale with Transfer Learning and Machine Teaching