Steering Large Language Models using Conceptors: Improving Addition-Based Activation Engineering

TL;DR

This paper introduces conceptors, a novel mathematical tool for more precise and effective activation control in large language models, outperforming traditional steering methods across various tasks.

Contribution

We propose conceptors as a new approach for activation engineering in LLMs, enabling more accurate and flexible control than existing methods.

Findings

Conceptors outperform traditional steering vectors in multiple tasks.

Boolean operations on conceptors improve combined steering goals.

Code implementation is publicly available on GitHub.

Abstract

Large language models have transformed AI, yet reliably controlling their outputs remains a challenge. This paper explores activation engineering, where outputs of pre-trained LLMs are controlled by manipulating their activations at inference time. Unlike traditional methods using a single steering vector, we introduce conceptors - mathematical constructs that represent sets of activation vectors as ellipsoidal regions. Conceptors act as soft projection matrices and offer more precise control over complex activation patterns. Our experiments demonstrate that conceptors outperform traditional methods across multiple steering tasks. We further use Boolean operations on conceptors for combined steering goals that empirically outperform additively combining steering vectors on a set of tasks. These results highlight conceptors as a promising tool for more effective steering of LLMs. Our code is available on github.com/jorispos/conceptorsteering.