Disentangled Safety Adapters Enable Efficient Guardrails and Flexible Inference-Time Alignment

TL;DR

Disentangled Safety Adapters (DSA) offer a modular approach to AI safety, enabling efficient, flexible safety functionalities and dynamic inference-time alignment adjustments without significant performance trade-offs.

Contribution

The paper introduces DSA, a novel framework that decouples safety mechanisms from the base model, improving safety, flexibility, and efficiency in AI systems.

Findings

DSA-based safety guardrails outperform standalone models in hate speech and hallucination detection.

DSA enables dynamic, inference-time safety alignment adjustments.

Combining safety guardrails and alignment with DSA significantly improves safety metrics with minimal performance loss.

Abstract

Existing paradigms for ensuring AI safety, such as guardrail models and alignment training, often compromise either inference efficiency or development flexibility. We introduce Disentangled Safety Adapters (DSA), a novel framework addressing these challenges by decoupling safety-specific computations from a task-optimized base model. DSA utilizes lightweight adapters that leverage the base model's internal representations, enabling diverse and flexible safety functionalities with minimal impact on inference cost. Empirically, DSA-based safety guardrails substantially outperform comparably sized standalone models across hate speech classification, detecting unsafe model inputs and responses, and hallucination detection with relative improvements of up to 53% in AUC. Furthermore, DSA-based safety alignment allows dynamic, inference-time adjustment of alignment strength and a fine-grained trade-off between instruction following performance and model safety. Importantly, combining the DSA safety guardrail with DSA safety alignment facilitates context-dependent alignment strength, boosting safety on StrongREJECT by 93% while maintaining 98% performance on MTBench - a total reduction in alignment tax of 8 percentage points compared to standard safety alignment fine-tuning. Overall, DSA presents a promising path towards more modular, efficient, and adaptable AI safety and alignment.