Scaling laws for activation steering with Llama 2 models and refusal mechanisms

2025-07-15

AI Summary

This work investigates the scalability of Contrastive Activation Addition (CAA) across the Llama 2 family (7B/13B/70B), focusing on its diminishing efficacy in steering refusal behavior. Methodologically, we intervene in the residual stream via direction vectors derived from contrastive token pairs, systematically evaluating the impact of model scale, transformer layer position, and guidance direction (positive vs. negative). Results reveal: (i) CAA effectiveness degrades significantly with increasing model size; (ii) negative-direction guidance exhibits greater robustness than positive-direction guidance; and (iii) optimal intervention layers concentrate in the mid-to-early transformer blocks. To our knowledge, this is the first study to empirically uncover scaling laws and layer-wise biases in activation-based steering—providing foundational evidence and design insights for scalable alignment techniques in large language models.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models
• Natural Language Processing: (Large) Language Models

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

As large language models (LLMs) evolve in complexity and capability, the efficacy of less widely deployed alignment techniques are uncertain. Building on previous work on activation steering and contrastive activation addition (CAA), this paper explores the effectiveness of CAA with model scale using the family of Llama 2 models (7B, 13B, and 70B). CAA works by finding desirable 'directions' in the model's residual stream vector space using contrastive pairs (for example, hate to love) and adding this direction to the residual stream during the forward pass. It directly manipulates the residual stream and aims to extract features from language models to better control their outputs. Using answer matching questions centered around the refusal behavior, we found that 1) CAA is most effective when applied at early-mid layers. 2) The effectiveness of CAA diminishes with model size. 3) Negative steering has more pronounced effects than positive steering across all model sizes.

Problem

• Effectiveness of activation steering with Llama 2 models
• Impact of model scale on contrastive activation addition
• Comparison of positive and negative steering effects

Innovation

• Uses contrastive activation addition (CAA) technique
• Manipulates residual stream vector space
• Applies CAA at early-mid layers