Steering Large Language Model Activations in Sparse Spaces

2025-02-28

AI Summary

This work addresses the challenge of behavior steering during the inference phase of large language models (LLMs), where entangled activations in dense representation spaces hinder both interpretability and precise behavioral control. To tackle this, we propose Sparse Activation Steering (SAS), a method that integrates sparse autoencoders (SAEs) with contrastive prompt pairs to identify and intervene upon monosemantic behavioral features within the sparse activation space—enabling interpretable and editable behavior modulation. Experiments on Gemma-2 demonstrate SAS’s capability for fine-grained behavioral control; moreover, scaling SAE size significantly improves the monosemanticity of SAS steering vectors, enhancing intervention reliability and mechanistic interpretability. To our knowledge, this is the first work to systematically incorporate sparse representation learning into test-time alignment steering for LLMs, establishing a novel paradigm for AI alignment.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Intelligent Robots: Behavior Learning & Control
• Cognitive Modeling & Cognitive Systems: Adaptive Behavior

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

A key challenge in AI alignment is guiding large language models (LLMs) to follow desired behaviors at test time. Activation steering, which modifies internal model activations during inference, offers a potential solution. However, prior work in dense activation spaces struggles with superposition, wherein multiple features become entangled, limiting interpretability and precise control. In contrast, sparse representations provide an untapped opportunity for more interpretable behavior modulation. In this work, we introduce sparse activation steering (SAS), a method that leverages sparse autoencoders (SAEs) to steer LLM behavior in sparse spaces. By isolating behavior-specific features through a contrastive prompt-pairing approach, we define a set of features that can selectively reinforce or suppress behaviors. Experiments on Gemma 2 LLMs show that SAS vectors enable nuanced behavioral modulation and finer-grained control. Furthermore, scaling SAEs improves monosemanticity of SAS vectors, suggesting more reliable and interpretable interventions.

Problem

• Guiding large language models to follow desired behaviors
• Overcoming superposition in dense activation spaces
• Enhancing interpretability and control with sparse activation steering

Innovation

• Sparse activation steering for LLM behavior control
• Contrastive prompt-pairing isolates behavior-specific features
• Scaling SAEs enhances interpretable behavioral interventions