Safety-Utility Conflicts Are Not Global: Surgical Alignment via Head-Level Diagnosis
This addresses the safety-utility trade-off in LLMs for AI safety applications, offering an interpretable and parameter-efficient approach, though it is incremental as it builds on existing mitigation strategies by focusing on head-level heterogeneity.
The paper tackled the problem of safety alignment in Large Language Models causing unintended degradation of general capabilities by showing that conflicts are not uniformly distributed across attention heads, and found that skipping a small group of high-conflict heads during training significantly reduces this loss without compromising safety.
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.