Steer2Adapt: Dynamically Composing Steering Vectors Elicits Efficient Adaptation of LLMs

Pengrui Han, Xueqiang Xu, Keyang Xuan, Peiyang Song, Siru Ouyang, Runchu Tian, Yuqing Jiang, Cheng Qian, Pengcheng Jiang, Jiashuo Sun, Junxia Cui, Ming Zhong, Ge Liu, Jiawei Han, Jiaxuan You · 2026

Read Steer2Adapt: Dynamically Composing Steering Vectors Elicits Efficient Adaptation of LLMs by Pengrui Han, Xueqiang Xu, Keyang Xuan, Peiyang Song, Siru Ouyang, Runchu Tian, Yuqing Jiang, Cheng Qian, Pengcheng Jiang, Jiashuo Sun, Junxia Cui, Ming Zhong, Ge Liu, Jiawei Han, Jiaxuan You (2026) 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-2602-07276.txt.

Description

Steering a language model with activation vectors is cheap, but one fixed direction per task cannot handle tasks that mix several capabilities. Pengrui Han and fourteen co-authors present Steer2Adapt, which composes steering vectors instead of learning each one from scratch.

Tasks are assumed to share a small set of underlying concept dimensions, captured as a reusable low-dimensional subspace. A new task is handled by finding a combination of basis vectors from only a handful of cases. Experiments across nine tasks and three models in thinking and safety show the approach working. Users get fast adaptation that reuses structure instead of relearning it each time.

Catalog record

CREATORPengrui Han, Xueqiang Xu, Keyang Xuan, Peiyang Song, Siru Ouyang, Runchu Tian, Yuqing Jiang, Cheng Qian, Pengcheng Jiang, Jiashuo Sun, Junxia Cui, Ming Zhong, Ge Liu, Jiawei Han, Jiaxuan You
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-07276

Browse the full library · About the Tower of Babel