DEAL: Disentangling Transformer Head Activations for LLM Steering (2506.08359v1)

Abstract: Inference-time steering aims to alter the response characteristics of LLMs without modifying their underlying parameters. A critical step in this process is the identification of internal modules within LLMs that are associated with the target behavior. However, current approaches to module selection often depend on superficial cues or ad-hoc heuristics, which can result in suboptimal or unintended outcomes. In this work, we propose a principled causal-attribution framework for identifying behavior-relevant attention heads in transformers. For each head, we train a vector-quantized autoencoder (VQ-AE) on its attention activations, partitioning the latent space into behavior-relevant and behavior-irrelevant subspaces, each quantized with a shared learnable codebook. We assess the behavioral relevance of each head by quantifying the separability of VQ-AE encodings for behavior-aligned versus behavior-violating responses using a binary classification metric. This yields a behavioral relevance score that reflects each head discriminative capacity with respect to the target behavior, guiding both selection and importance weighting. Experiments on seven LLMs from two model families and five behavioral steering datasets demonstrate that our method enables more accurate inference-time interventions, achieving superior performance on the truthfulness-steering task. Furthermore, the heads selected by our approach exhibit strong zero-shot generalization in cross-domain truthfulness-steering scenarios.