Blog

10 December 2025

Generalization or Hallucination? Understanding Out-of-Context Reasoning in Transformers

NIPS'25

๐Ÿ’กGeneralization์ด๋“  Hallucination์ด๋“  ๋ชจ๋‘ ๋‹ค Out-of-Context Reasoning์˜ ํ˜„์ƒ์ด๊ณ , ์ด๋Š” Output ํ–‰๋ ฌ๊ณผ Value ํ–‰๋ ฌ์ด ๋ถ„๋ฆฌ๋˜์–ด์žˆ์–ด ํ•™์Šต๊ฐ€๋Šฅํ•˜๋‹ค!

์ด์Šนํ™˜
10 December 2025

Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models

ICLR'25

๐Ÿ’กLLM ์•ˆ์—๋Š” ์ด ์—”ํ‹ฐํ‹ฐ๋ฅผ LLM์ด ์•„๋Š”์ง€/๋ชจ๋ฅด๋Š”์ง€๋ฅผ ํ‘œ์‹œํ•˜๋Š” latent ๋ฐฉํ–ฅ์ด ์‹ค์ œ๋กœ ์กด์žฌ์ด latent ๋ฐฉํ–ฅ์„ ์กฐ์ž‘(steering) ํ•˜๋ฉด,์›๋ž˜๋Š” ๋ชจ๋ฅธ๋‹ค๊ณ  ๋งํ•˜๋˜ ์งˆ๋ฌธ(๋‹ต๋ณ€ ๊ฑฐ๋ถ€)์— ๋Œ€ํ•ด ํ• ๋ฃจ์‹œ๋„ค์ด์…˜์„ ์‹œํ‚ค๊ฑฐ๋‚˜,์›๋ž˜ ์ž˜ ์•Œ๋˜ ์—”ํ‹ฐํ‹ฐ์— ๋Œ€ํ•ด์„œ๋„ ๋‹ต๋ณ€์„ ๊ฑฐ๋ถ€ํ•˜๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Œ

26 November 2025

On the Role of Attention Heads in Large Language Model Safety

ICLR'25

๐Ÿ’กLLM ์•ˆ์ „์„ฑ์€ ์‚ฌ์‹ค ์†Œ์ˆ˜์˜ attention head ์— ์ง‘์ค‘๋˜์–ด ์žˆ์–ด์„œ, ๊ทธ head๋“ค๋งŒ ์‚ด์ง ๊บผ๋„ ๐Ÿšจ ์•ˆ์ •์„ฑ์ด ๋ฐ”๋กœ ๋ฌด๋„ˆ์ง„๋‹ค๋Š” ๊ฑธ ๋ฐํž˜ ๐Ÿ” ShipsยทSahara๋กœ ์–ด๋–ค head๊ฐ€ ์ง„์งœ safety ๋‹ด๋‹น์ธ์ง€ ์ฐพ์•„๋‚ด๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•จ โš™๏ธ๐Ÿ”ฅ