arXiv cs.CL· Anna C. Marbut, Daniel R. Olson, Travis J. Wheeler·· 12 小时前AI 评分31
在 MoE 模型中分解拒绝引导机制
Decomposing Refusal Steering in Mixture-of-Experts Models
AI 导读
研究将拒绝引导（refusal steering）方法扩展到三个开源 MoE 大模型，发现引导效果不受 MoE 复杂路由模式影响，单个专家在可自选位置时平均可复现 78% 的完整引导效果。但若将专家限制在全层引导所用的位置，复现率降至平均 54%，在安全相关系统提示下差距进一步扩大。拒绝信号与专家路由行为所捕获的信号并不相同，表明拒绝行为在 MoE 架构中分布不均，而非由单一统一机制控制。
正文 · 原文
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.
| Subjects: | Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2606.04160 [cs.CL] |
| (or arXiv:2606.04160v3 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2606.04160 arXiv-issued DOI via DataCite |
Submission history
From: Anna Marbut [view email]
[v1]
Tue, 2 Jun 2026 19:23:46 UTC (237 KB)
[v2]
Sat, 19 Sep 2026 21:15:06 UTC (248 KB)
[v3]
Tue, 20 Oct 2026 16:57:45 UTC (248 KB)
来源：arXiv cs.CL · arxiv.org