Disentangled Safety Adapters Enable Efficient Guardrails and Flexible Inference-Time Alignment
Kundan Krishna, Joseph Y Cheng, Charles Maalouf, Leon A Gatys
Why It Matters
What makes this one worth your time
AI engineers and researchers should care because DSA provides a method to implement safety measures without significantly impacting performance, offering a balance between safety and task efficiency.
DSA offers a modular approach to AI safety, enhancing efficiency and flexibility in safety and alignment tasks.
Summary
The paper introduces Disentangled Safety Adapters (DSA), a framework that separates safety computations from task-optimized models, allowing for efficient and flexible safety functionalities with minimal inference cost. The framework is shown to improve performance in safety tasks such as hate speech classification and hallucination detection, and allows for dynamic adjustment of alignment strength during inference.
Key contributions
- Introduction of Disentangled Safety Adapters (DSA) for AI safety.
- Empirical evidence showing improved performance in safety tasks with DSA.
- Dynamic adjustment of alignment strength at inference time.
Notable insights
- Decoupling safety computations from task models can enhance flexibility and efficiency.
- Dynamic inference-time adjustment of alignment strength allows for context-dependent safety measures.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2506.00166v2 Announce Type: replace-cross 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.