Adaptive Task Vectors for Large Language Models
Joonseong KangSoojeong LeeSubeen ParkSumin Park Taero KimJihee KimRyunyi Lee, Kyungwoo Song Yonsei University · 2025
Read Adaptive Task Vectors for Large Language Models by Joonseong KangSoojeong LeeSubeen ParkSumin Park Taero KimJihee KimRyunyi Lee, Kyungwoo Song Yonsei University (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-03426.txt.
Description
Joonseong Kang and co-authors make in 2025 a smarter shortcut for teaching language models new tasks. Showing examples in the prompt works but is slow and fussy. Compressing the lesson into one vector is faster, yet a fixed vector may not fit each new question.
Their method, Adaptive Task Vectors, uses a small model to build a fresh vector for every input. That vector is reshaped to fit the big model and steers its answer. Tests show strong results even on tasks never seen in training. Theory adds support: at equal budget the method matches a popular fine-tuning technique in power and beats another, prefix tuning.
Catalog record
| CREATOR | Joonseong KangSoojeong LeeSubeen ParkSumin Park Taero KimJihee KimRyunyi Lee, Kyungwoo Song Yonsei University |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 67 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-03426