Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control

Julian Skifstad, Xinyue Annie Yang, Glen Chou · 2026

Read Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control by Julian Skifstad, Xinyue Annie Yang, Glen Chou (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-2604-19018.txt.

Description

Skifstad, Yang, and Chou show LLM layer dynamics are locally linear, enabling activation steering by linear optimal control. Current steering uses open loop interventions that ignore perturbation spread across layers and lack error feedback. They model inference as a linear time varying system and adapt the linear quadratic regulator, using layer Jacobians to steer activations to setpoints in closed loop. They also bound tracking error, giving formal steering guarantees. With an adaptive setpoint signal, the method controls toxicity, truthfulness, refusal, and arbitrary concepts, beating baselines.

Catalog record

CREATORJulian Skifstad, Xinyue Annie Yang, Glen Chou
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE108323
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.19018

Browse the full library · About the Tower of Babel