analiza aiqbench
This research investigates how refusal steering vectors function within Mixture-of-Experts (MoE) architectures, revealing that safety alignment is distributed unevenly across experts rather than governed by a single mechanism. The study demonstrates that while individual experts can recover most of the full-layer steering effect when unconstrained, their efficacy drops significantly when restricted to standard routing locations, particularly under safety prompts. These findings provide critical insights into the internal mechanics of aligning complex MoE models for reliable harm reduction.
Decomposing Refusal Steering in Mixture-of-Experts Models
arXiv:2606.04160v2 Announce Type: replace Abstract: Safety alignment in instruction-tuned large language models (LLMs) depends on a model's ability to reliably refuse harmful or disallowed requests. Recent work has shown that a steering vector can be applied to a dense LLM during inference to suppress refusal behavior and induce responses to harmful requests. We extend this refusal steering method to three open-source Mixture-of-Experts (MoE) LLMs to decompose and better understand how refusal mechanisms operate across MoE components. We find that steering performance is uninhibited by the complex routing patterns inherent to the MoE architecture, and that a single expert recovers 78% of the full steering effect on average when free to select its own location. However, expert-level steering only recovers 54% on average when constrained to the location used by full-layer steering, a gap that widens further under safety-related system prompts. Our results also show that refusal signals captured by these steering methods differ from those captured by expert routing behavior. Together, these results suggest that refusal behavior is distributed unevenly across the MoE LLM architecture rather than governed by a single unified mechanism.
Abstract:Safety alignment in instruction-tuned large language models (LLMs) depends on a model's ability to reliably refuse harmful or disallowed requests. Recent work has shown that a steering vector can be applied to a dense LLM during inference to suppress refusal behavior and induce responses to harmful requests. We extend this refusal steering method to three open-source Mixture-of-Experts (MoE) LLMs to decompose and better understand how refusal mechanisms operate across MoE components. We find that steering performance is uninhibited by the complex routing patterns inherent to the MoE architecture, and that a single expert recovers 78% of the full steering effect on average when free to select its own location. However, expert-level steering only recovers 54% on average when constrained to the location used by full-layer steering, a gap that widens further under safety-related system prompts. Our results also show that refusal signals captured by these steering methods differ from those captured by expert routing behavior. Together, these results suggest that refusal behavior is distributed unevenly across the MoE LLM architecture rather than governed by a single unified mechanism.
Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2606.04160 [cs.CL]
(or arXiv:2606.04160v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2606.04160 Focus to learn more arXiv-issued DOI via DataCite
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.
Text integral sub licența deschisă a publicației · 274 cuvinte
Subiecte conexe
Expert-Aware Refusal Steering
arXiv · 5 iun. 2026
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · 8 oct. 2026
Decided Upstream, Written Late: Locating and Pricing the Cross-Lingual Refusal Circuit of a Multilingual MoE
arXiv · 19 aug. 2026
Reading, Not Manipulating: Leveraging Router Logits for Multimodal Safety in MoE Vision-Language Models
arXiv · 7 oct. 2026