aiqbench analysis
This research demonstrates that refusal steering techniques, previously effective on dense models, remain robust when applied to Mixture-of-Experts (MoE) architectures despite their complex routing dynamics. By leveraging expert-specific steering directions, the study reveals that attention mechanisms play a substantial role in MoE refusal behavior, distinct from standard routing patterns. These findings offer a practical pathway for maintaining safety alignment in open-source MoE models without compromising their architectural efficiency.
Expert-Aware Refusal Steering
arXiv:2606.04160v1 Announce Type: cross Abstract: Safety alignment in instruction-tuned large language models (LLMs) depends on a model's ability to reliably refuse to respond to harmful or disallowed requests. Recent work has shown that a steering vector can be applied to a dense LLM during inference to effectively suppress refusal behavior, inducing response to harmful requests. We extend this refusal steering method to three open-source Mixture-of-Experts (MoE) LLMs and find that steering performance is uninhibited by the complex routing patterns inherent to the MoE architecture. We then propose two expert-aware refusal steering methods that leverage refusal-specific expert routing patterns and expert-specific steering directions to suppress normal refusal behavior. We find that refusal behavior can be effectively steered based on the output of a single expert. Our results show that refusal signals captured by steering methods differ from expert routing behavior, suggesting a substantial role for attention in MoE refusal behavior.
Abstract:Safety alignment in instruction-tuned large language models (LLMs) depends on a model's ability to reliably refuse to respond to harmful or disallowed requests. Recent work has shown that a steering vector can be applied to a dense LLM during inference to effectively suppress refusal behavior, inducing response to harmful requests. We extend this refusal steering method to three open-source Mixture-of-Experts (MoE) LLMs and find that steering performance is uninhibited by the complex routing patterns inherent to the MoE architecture. We then propose two expert-aware refusal steering methods that leverage refusal-specific expert routing patterns and expert-specific steering directions to suppress normal refusal behavior. We find that refusal behavior can be effectively steered based on the output of a single expert. Our results show that refusal signals captured by steering methods differ from expert routing behavior, suggesting a substantial role for attention in MoE refusal behavior.
Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2606.04160 [cs.CL]
(or arXiv:2606.04160v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2606.04160 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 242 words
Related coverage
"About a quarter of them contained bugs": Gemini 4's ML R&D benchmark
LessWrong · Oct 16, 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · Sep 18, 2026
RASET: Router-Agnostic Safety-Critical Expert Tuning Exposes Localized Safety Enforcement Failures in Mixture-of-Experts LLMs
arXiv · Aug 25, 2026
The Continual Constitutional Convention: How language models revise the documents that govern them
AI Alignment Forum · Oct 13, 2026