Safety-Utility Conflicts Are Not Global: Surgical Alignment via Head-Level Diagnosis

TL;DR

This paper introduces Conflict-Aware Sparse Tuning (CAST), a head-level diagnosis method for LLM safety alignment that selectively updates parameters, reducing safety-utility conflicts and preserving capabilities.

Contribution

It proposes a novel framework that diagnoses and selectively updates attention heads in transformers, addressing the limitations of global gradient-based methods.

Findings

Alignment conflicts are unevenly distributed across heads.

Skipping high-conflict heads preserves capabilities while improving safety.

Selective head tuning enhances safety-utility trade-offs.

Abstract

Safety alignment in Large Language Models (LLMs) inherently presents a multi-objective optimization conflict, often accompanied by an unintended degradation of general capabilities. Existing mitigation strategies typically rely on global gradient geometry to resolve these conflicts, yet they overlook Modular Heterogeneity within Transformers, specifically that the functional sensitivity and degree of conflict vary substantially across different attention heads. Such global approaches impose uniform update rules across all parameters, often resulting in suboptimal trade-offs by indiscriminately updating utility sensitive heads that exhibit intense gradient conflicts. To address this limitation, we propose Conflict-Aware Sparse Tuning (CAST), a framework that integrates head-level diagnosis with sparse fine-tuning. CAST first constructs a pre-alignment conflict map by synthesizing Optimization Conflict and Functional Sensitivity, which then guides the selective update of parameters. Experiments reveal that alignment conflicts in LLMs are not uniformly distributed. We find that the drop in general capabilities mainly comes from updating a small group of ``high-conflict'' heads. By simply skipping these heads during training, we significantly reduce this loss without compromising safety, offering an interpretable and parameter-efficient approach to improving the safety-utility trade-off.