Paper Detail
Brandon Hsu, Daniel Beaglehole, Adityanarayanan Radhakrishnan, Mikhail Belkin
Linear activation steering is a powerful approach for eliciting the capabilities of large language models and specializing their behavior using limited labeled data. While effective, existing methods often apply a fixed steering strength to all tokens, resulting in inconsistent steering quality across diverse input prompts. In this work, we introduce Contextual Linear Activation Steering (CLAS), a method that dynamically adapts linear activation steering to context-dependent steering strengths. Across eleven steering benchmarks and four model families, it consistently outperforms standard linear activation steering and matches or exceeds the performance of ReFT and LoRA in settings with limited labeled data. We therefore propose CLAS as a scalable, interpretable, and accurate method for specializing and steering large language models.
No structured notes yet. Add `summary_sections`, `why_relevant`, `claim_impact`, or `next_action` in `papers.jsonl` to enrich this view.
No ranking explanation is available yet.
No tags.
@article{hsu2026contextual,
title = {Contextual Linear Activation Steering of Language Models},
author = {Brandon Hsu and Daniel Beaglehole and Adityanarayanan Radhakrishnan and Mikhail Belkin},
year = {2026},
abstract = {Linear activation steering is a powerful approach for eliciting the capabilities of large language models and specializing their behavior using limited labeled data. While effective, existing methods often apply a fixed steering strength to all tokens, resulting in inconsistent steering quality across diverse input prompts. In this work, we introduce Contextual Linear Activation Steering (CLAS), a method that dynamically adapts linear activation steering to context-dependent steering strengths. },
url = {https://arxiv.org/abs/2604.24693},
keywords = {cs.CL},
eprint = {2604.24693},
archiveprefix = {arXiv},
}
{}