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

2024-10-09

arXiv.org

AI Summary

This work addresses the low precision and poor interpretability of output control during large language model (LLM) inference. We propose a novel activation-space intervention method based on Conceptors—the first application of Conceptors to LLM control. Our approach models semantic concepts as ellipsoidal regions in the activation space and employs soft projection matrices to achieve nonlinear, robust activation modulation. Crucially, it supports Boolean operations (e.g., AND, OR), enabling compositional, cooperative concept guidance and overcoming the linearity limitations of conventional vector arithmetic. Experiments across multiple controllable generation tasks demonstrate consistent superiority over baselines: Boolean-combined Conceptors yield an average 3.2% absolute accuracy improvement over simple vector addition. The implementation is publicly available.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Natural Language Processing: (Large) Language Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

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.

Problem

• Control
• Large Language Models
• Output Precision

Innovation

• Conceptual Bodies
• Activation Engineering
• Complex Steering