Current paradigms for guaranteeing AI security, comparable to guardrail fashions and alignment coaching, usually compromise both inference effectivity or improvement flexibility. We introduce Disentangled Security Adapters (DSA), a novel framework addressing these challenges by decoupling safety-specific computations from a task-optimized base mannequin. DSA makes use of light-weight adapters that leverage the bottom mannequin’s inner representations, enabling numerous and versatile security functionalities with minimal impression on inference value. Empirically, DSA-based security guardrails considerably outperform comparably sized standalone fashions, notably bettering hallucination detection (0.88 vs. 0.61 AUC on Summedits) and in addition excelling at classifying hate speech (0.98 vs. 0.92 on ToxiGen) and unsafe mannequin inputs and responses (0.93 vs. 0.90 on AEGIS2.0 & BeaverTails). Moreover, DSA-based security alignment permits dynamic, inference-time adjustment of alignment energy and a fine-grained trade-off between instruction following efficiency and mannequin security. Importantly, combining the DSA security guardrail with DSA security alignment facilitates context-dependent alignment energy, boosting security on StrongReject by 93% whereas sustaining 98% efficiency on MTBench — a complete discount in alignment tax of 8 proportion factors in comparison with normal security alignment fine-tuning. Total, DSA presents a promising path in direction of extra modular, environment friendly, and adaptable AI security and alignment.
Determine 1: Overview of DSA structure and the way it compares to plain security methods.
