Understanding On-Policy Distillation: A Mechanistic Interpretability Perspective via Sparse Crosscoders
The swap readout is proposed, which reads each student checkpoint's feature activations on its own, measuring how training changes the student's use of each feature, even for checkpoints unseen by the crosscoder.
Zi-Chao Yu, Qian-Shuo Ye, Xu Wang et al.
· 0 citations