HyperSteer: Activation Steering at Scale with Hypernetworks

2025-06-03

AI Summary

Existing activation steering methods face a fundamental trade-off: unsupervised dictionary learning scales well but lacks guarantees on individual vector efficacy and task coverage; supervised approaches achieve high precision yet require costly, per-vector annotation and retraining. This paper introduces the first end-to-end hypernetwork-based activation steering framework, which dynamically generates high-quality steering vectors conditioned on natural-language steering prompts and internal language model activations. By innovatively integrating hypernetworks into activation steering, our method enables zero-shot generalization—unifying the precision of supervised methods with the scalability of unsupervised ones. Evaluated across over one thousand diverse prompts, it significantly outperforms prior state-of-the-art approaches. Crucially, it exhibits strong generalization to unseen prompts, achieving performance on par with prompt-based steering while operating entirely in the activation space.

Technology Category

• Machine Learning: Active Learning
• Natural Language Processing: Generation
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Large pretrained models with web data
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Steering language models (LMs) by modifying internal activations is a popular approach for controlling text generation. Unsupervised dictionary learning methods, e.g., sparse autoencoders, can be scaled to produce many steering vectors, but lack guarantees on the individual efficacy of each vector and control over the coverage of relevant steering tasks. In contrast, supervised methods for constructing steering vectors are targeted and effective, but require more data collection and training for each additional steering vector produced. In this work, we introduce HyperSteer, a family of hypernetwork-based architectures which are trained end-to-end to generate steering vectors conditioned on the natural language steering prompts and the internals of the steered LM. In our evaluations, we show that scaling HyperSteer with thousands of steering prompts exceeds the performance of state-of-the-art activation steering methods, even on steering prompts never seen during training. Moreover, HyperSteer performs on par with steering-via-prompting.

Problem

• Lack of guaranteed efficacy in unsupervised steering vectors
• Supervised steering vectors require excessive data and training
• Need for scalable, effective activation steering in language models

Innovation

• Hypernetwork-based architectures for steering vectors
• Conditioned on natural language steering prompts
• Scales with thousands of steering prompts effectively