作者位次 #2/8ACL 2026Poster· Natural Language Processing
Mechanistic Interpretability Should Prioritize Feature Consistency in Sparse Autoencoders
Xiangchen Song · Aashiq Muhamed · Yujia Zheng · Lingjing Kong · Zeyu Tang · Mona T. Diab · Virginia Smith · Kun Zhang
摘要
Sparse Autoencoders (SAEs) are a prominent tool in mechanistic interpretability (MI) for decomposing neural network activations into interpretable features. However, the aspiration to identify a canonical set of features is challenged by the observed inconsistency of learned SAE features across d...