Controlling Large Language Models Through Concept Activation Vectors

2025-01-10

AI Summary

Existing LLM controllable generation methods suffer from coarse-grained control, high computational overhead, or reliance on fine-tuning—compromising accuracy, efficiency, and interpretability simultaneously. To address this, we propose GCAV, a novel plug-and-play intervention framework grounded in Concept Activation Vectors (CAVs). GCAV learns interpretable, semantic directions—e.g., toxicity, sentiment, or writing style—within latent representation spaces, enabling layer-wise, sample-specific, and intensity-controllable inference-time steering. Crucially, it requires no model modification or fine-tuning, only lightweight linear projection and directional offset operations. Extensive experiments demonstrate that GCAV achieves state-of-the-art performance across diverse controllable generation tasks: reducing toxicity by 42%, significantly improving precision in sentiment polarity and stylistic control, and incurring negligible computational overhead.

Technology Category

• Natural Language Processing: Generation
• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Machine Learning: Large Multimodal Models (LMMs)

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

As large language models (LLMs) are widely deployed across various domains, the ability to control their generated outputs has become more critical. This control involves aligning LLMs outputs with human values and ethical principles or customizing LLMs on specific topics or styles for individual users. Existing controlled generation methods either require significant computational resources and extensive trial-and-error or provide coarse-grained control. In this paper, we propose Generation with Concept Activation Vector (GCAV), a lightweight model control framework that ensures accurate control without requiring resource-extensive fine-tuning. Specifically, GCAV first trains a concept activation vector for specified concepts to be controlled, such as toxicity. During inference, GCAV steers the concept vector in LLMs, for example, by removing the toxicity concept vector from the activation layers. Control experiments from different perspectives, including toxicity reduction, sentiment control, linguistic style, and topic control, demonstrate that our framework achieves state-of-the-art performance with granular control, allowing for fine-grained adjustments of both the steering layers and the steering magnitudes for individual samples.

Problem

• Large Language Models
• Content Control
• Resource Efficiency

Innovation

• GCAV
• Controlled Language Generation
• Lightweight Method