Channel Location Constrains the Auditability of Subliminal Learning

The paper introduces the concept of channel location as a critical factor in the auditability of subliminal learning, where a student model inherits traits from a teacher model through distillation data. It identifies three regimes of trait transfer, emphasizing that the effectiveness of audits depends on the initialization and channel through which traits are conveyed, with metrics such as coverage and cosine similarity demonstrating high predictive power for transfer success. This research is significant for practitioners as it highlights the limitations of current auditing methods and the potential for unintended trait transfer in pretrained language models, suggesting that naive removal of training data does not guarantee the elimination of learned biases.