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

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Skifstad, Julian, Yang, Xinyue Annie, Chou, Glen
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866911611124449280
author Skifstad, Julian
Yang, Xinyue Annie
Chou, Glen
author_facet Skifstad, Julian
Yang, Xinyue Annie
Chou, Glen
contents Inference-time LLM alignment methods, particularly activation steering, offer an alternative to fine-tuning by directly modifying activations during generation. Existing methods, however, often rely on non-anticipative interventions that ignore how perturbations propagate through transformer layers and lack online error feedback, resulting in suboptimal, open-loop control. To address this, we show empirically that, despite the nonlinear structure of transformer blocks, layer-wise dynamics across multiple LLM architectures and scales are well-approximated by locally-linear models. Exploiting this property, we model LLM inference as a linear time-varying dynamical system and adapt the classical linear quadratic regulator to compute feedback controllers using layer-wise Jacobians, steering activations toward desired semantic setpoints in closed-loop with minimal computational overhead and no offline training. We also derive theoretical bounds on setpoint tracking error, enabling formal guarantees on steering performance. Using a novel adaptive semantic feature setpoint signal, our method yields robust, fine-grained behavior control across models, scales, and tasks, including state-of-the-art modulation of toxicity, truthfulness, refusal, and arbitrary concepts, surpassing baseline steering methods. Our code is available at: https://github.com/trustworthyrobotics/lqr-activation-steering
format Preprint
id arxiv_https___arxiv_org_abs_2604_19018
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control
Skifstad, Julian
Yang, Xinyue Annie
Chou, Glen
Machine Learning
Artificial Intelligence
Systems and Control
Optimization and Control
Inference-time LLM alignment methods, particularly activation steering, offer an alternative to fine-tuning by directly modifying activations during generation. Existing methods, however, often rely on non-anticipative interventions that ignore how perturbations propagate through transformer layers and lack online error feedback, resulting in suboptimal, open-loop control. To address this, we show empirically that, despite the nonlinear structure of transformer blocks, layer-wise dynamics across multiple LLM architectures and scales are well-approximated by locally-linear models. Exploiting this property, we model LLM inference as a linear time-varying dynamical system and adapt the classical linear quadratic regulator to compute feedback controllers using layer-wise Jacobians, steering activations toward desired semantic setpoints in closed-loop with minimal computational overhead and no offline training. We also derive theoretical bounds on setpoint tracking error, enabling formal guarantees on steering performance. Using a novel adaptive semantic feature setpoint signal, our method yields robust, fine-grained behavior control across models, scales, and tasks, including state-of-the-art modulation of toxicity, truthfulness, refusal, and arbitrary concepts, surpassing baseline steering methods. Our code is available at: https://github.com/trustworthyrobotics/lqr-activation-steering
title Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control
topic Machine Learning
Artificial Intelligence
Systems and Control
Optimization and Control
url https://arxiv.org/abs/2604.19018