Anwar et al. show CoT monitorability recovers under an unsupervised post-training objective
Mutual information between chain-of-thought and output can become a training stage, with a stake for labs that already compress or pressure their traces.
Read the piece