Why Ghost Outputs Teach: A Kernel-Based Understanding of Subliminal Learning
A chained cross-task kernel is derived that explicitly links ghost-output supervision to changes in task predictions through shared backbone representations, providing the first learning-dynamics-based theoretical explanation of how ghost-output supervision gives rise to subliminal learning.
Zhe Li, B. Ying, Chao-Sheng Dong et al.
· 0 citations