作者位次 #2/9ACL 2026Findings· Findings of ACL
Safe- SAIL : Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework
Jiaqi Weng · Han Zheng · Hanyu Zhang · Ej Zhou · Qinqin He · Jialing Tao · Hui Xue · Zhixuan Chu · Xiting Wang
摘要
Sparse autoencoders (SAEs) enable interpretability research by decomposing entangled model activations into monosemantic features. However, under what circumstances SAEs derive most fine-grained latent features for safety—a low-frequency concept domain—remains unexplored. Two key challenges exist...