Saved in:
Bibliographic Details
Main Authors: Krishna, Kundan, Cheng, Joseph Y, Maalouf, Charles, Gatys, Leon A
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2506.00166
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914522346815488
author Krishna, Kundan
Cheng, Joseph Y
Maalouf, Charles
Gatys, Leon A
author_facet Krishna, Kundan
Cheng, Joseph Y
Maalouf, Charles
Gatys, Leon A
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2506_00166
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Disentangled Safety Adapters Enable Efficient Guardrails and Flexible Inference-Time Alignment
Krishna, Kundan
Cheng, Joseph Y
Maalouf, Charles
Gatys, Leon A
Machine Learning
Artificial Intelligence
Computation and Language
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.
title Disentangled Safety Adapters Enable Efficient Guardrails and Flexible Inference-Time Alignment
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2506.00166