Fine-Grained Activation Steering: Steering Less, Achieving More Corresponding author.

Zijian Feng, Tianjiao Li, Zixiao Zhu, Hanzhang Zhou, Junlang Qian, Li Zhang, Jia Jim Deryl Chua, Lee Onn Mak, Gee Wah Ng, Kezhi Mao · 2026

Read Fine-Grained Activation Steering: Steering Less, Achieving More Corresponding author. by Zijian Feng, Tianjiao Li, Zixiao Zhu, Hanzhang Zhou, Junlang Qian, Li Zhang, Jia Jim Deryl Chua, Lee Onn Mak, Gee Wah Ng, Kezhi Mao (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-2602-04428.txt.

Description

Steering a language model by whole layers moves helpful and harmful features together. Zijian Feng and colleagues propose fine-grained activation steering in a 2026 paper. Current methods intervene on blocks: attention heads, feedforward networks, or residual streams bundled as units.

Splitting blocks into atomic units, single dimensions tied to slices of weight matrices, exposes the tangle. Different atomic units control distinct output token patterns, so block steering is bound to be coarse and intrusive. Steering single units promises the same change in behavior with less collateral damage. The analysis recasts activation steering as a precision instrument problem: steer less, achieve more, exactly as their title claims.

Catalog record

CREATORZijian Feng, Tianjiao Li, Zixiao Zhu, Hanzhang Zhou, Junlang Qian, Li Zhang, Jia Jim Deryl Chua, Lee Onn Mak, Gee Wah Ng, Kezhi Mao
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE57 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-04428

Browse the full library · About the Tower of Babel