Post

What are the next big trends in LLM research?

๐Ÿ’ก The LLM space is experiencing rapid progress, with new papers or releases almost every day.

If you aim to stay updated with the latest advancements, hereโ€™s a guide on emerging patterns: https://github.com/aishwaryanr/awesome-generative-ai-guide/blob/main/free_courses/Applied_LLMs_Mastery_2024/week10_research_trends.md

They are:

๐Ÿš€Multi-Modal LLMs

  • ๐Ÿ“•Combine text processing with multimodal components like audio, imagery and videos. Examples: OpenAI Sora, Gemini, LLaVA

๐Ÿš€Open-Source LLMs

  • ๐Ÿ“•Open-source models provide model weights, and optionally, checkpoints and training data, promoting fairness and transparency. Examples: LLM360, LLaMA, OLMo, Llama-3

๐Ÿš€Domain Specific LLMs

  • ๐Ÿ“•Domain-specific LLMs are tailored to excel in particular fields for example- code generation or biology, optimizing their performance accordingly. Examples: BioGPT, StarCoder, MathVista

๐Ÿš€LLM Agents

  • ๐Ÿ“•LLM agents are applications that LLMs combined with modules like planning and memory, to execute complex tasks. Examples: ChemCrow, ToolLLM, OS-Copilot

๐Ÿš€Smaller LLMs (Including Quantized LLMs)

  • ๐Ÿ“•LLMs with reduced precision or lesser parameters ideal for deployment on devices with limited resources. Examples: BitNet, Gemma 1B, Lit-LLaMA

๐Ÿš€Non-Transformer LLMs

  • ๐Ÿ“•LLMs that deviate from the standard transformer architecture (for example: incorporating RNNs) and offer solutions to transformer pain-points. Examples: Mamba, RMKV

 LLM Research Trends

Translate to Korean

๋‚ด ๊ฐ€์ด๋“œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค๊ฐ€์˜ค๋Š” ๋ชจ๋“  ํŠธ๋ Œ๋“œ๋ฅผ ๋”ฐ๋ผ์žก์œผ์„ธ์š”!

๐Ÿ’ก LLM ๋ถ„์•ผ๋Š” ๊ฑฐ์˜ ๋งค์ผ ์ƒˆ๋กœ์šด ๋…ผ๋ฌธ์ด๋‚˜ ๋ฐœํ‘œ๋ฅผ ํ†ตํ•ด ๊ธ‰์†ํ•œ ๋ฐœ์ „์„ ์ด๋ฃจ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์ตœ์‹  ๋ฐœ์ „ ์‚ฌํ•ญ์„ ์ตœ์‹  ์ƒํƒœ๋กœ ์œ ์ง€ํ•˜๋ ค๋Š” ๊ฒฝ์šฐ ์ƒˆ๋กœ์šด ํŒจํ„ด์— ๋Œ€ํ•œ ๊ฐ€์ด๋“œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. https://github.com/aishwaryanr/awesome-generative-ai-guide/blob/main/free_courses/Applied_LLMs_Mastery_2024/week10_research_trends.md

๊ทธ๋“ค์€:

๐Ÿš€๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ LLM

  • ๐Ÿ“•ํ…์ŠคํŠธ ์ฒ˜๋ฆฌ๋ฅผ ์˜ค๋””์˜ค, ์ด๋ฏธ์ง€ ๋ฐ ๋น„๋””์˜ค์™€ ๊ฐ™์€ ๋‹ค์ค‘ ๋ชจ๋“œ ๊ตฌ์„ฑ ์š”์†Œ์™€ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ: OpenAI Sora, Gemini, LLaVA

๐Ÿš€์˜คํ”ˆ ์†Œ์Šค LLM

  • ๐Ÿ“•์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์€ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์™€ ์„ ํƒ์ ์œผ๋กœ ์ฒดํฌํฌ์ธํŠธ ๋ฐ ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ•˜์—ฌ ๊ณต์ •์„ฑ๊ณผ ํˆฌ๋ช…์„ฑ์„ ์ด‰์ง„ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ: LLM360, LLaMA, OLMo, Llama-3

๐Ÿš€๋„๋ฉ”์ธ๋ณ„ LLM

  • ๐Ÿ“•๋„๋ฉ”์ธ๋ณ„ LLM์€ ์ฝ”๋“œ ์ƒ์„ฑ ๋˜๋Š” ์ƒ๋ฌผํ•™๊ณผ ๊ฐ™์€ ํŠน์ • ๋ถ„์•ผ์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋„๋ก ๋งž์ถคํ™”๋˜์–ด ๊ทธ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์„ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ: BioGPT, StarCoder, MathVista

๐Ÿš€LLM ์—์ด์ „ํŠธ

  • ๐Ÿ“•LLM ์—์ด์ „ํŠธ๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ์‹คํ–‰ํ•˜๊ธฐ ์œ„ํ•ด LLM์ด ๊ณ„ํš ๋ฐ ๋ฉ”๋ชจ๋ฆฌ์™€ ๊ฐ™์€ ๋ชจ๋“ˆ๊ณผ ๊ฒฐํ•ฉํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์ž…๋‹ˆ๋‹ค. ์˜ˆ: ChemCrow, ToolLLM, OS-Copilot

๐Ÿš€๋” ์ž‘์€ LLM(์–‘์žํ™”๋œ LLM ํฌํ•จ)

  • ๐Ÿ“•์ •๋ฐ€๋„๊ฐ€ ๋‚ฎ๊ฑฐ๋‚˜ ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ ์ ์€ LLM์€ ๋ฆฌ์†Œ์Šค๊ฐ€ ์ œํ•œ๋œ ์žฅ์น˜์— ๋ฐฐํฌํ•˜๋Š” ๋ฐ ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ: BitNet, Gemma 1B, Lit-LLaMA

๐Ÿš€๋น„๋ณ€์••๊ธฐ LLM

-๐Ÿ“•ํ‘œ์ค€ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜(์˜ˆ: RNN ํ†ตํ•ฉ)์—์„œ ๋ฒ—์–ด๋‚˜ ํŠธ๋žœ์Šคํฌ๋จธ ๋ฌธ์ œ์ ์— ๋Œ€ํ•œ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•˜๋Š” LLM์ž…๋‹ˆ๋‹ค. ์˜ˆ: ๋ง˜๋ฐ”, RMKV โ€”

This post is licensed under CC BY 4.0 by the author.