Fast Multi-dimensional Refusal Subspaces via RFM-AGOP
通过RFM-AGOP的快速多维拒绝子空间
AI总结 提出利用RFM-AGOP算法快速提取大型语言模型中的多维拒绝子空间,在推理和非推理模型上均实现秒级识别,且性能优于现有方法。
Comments Accepted to the Mechanistic Interpretability Workshop at the 43rd International Conference on Machine Learning, Seoul, South Korea, 2026