NLAs Miss Some Internalized Behaviors

NLAs (Natural Language Autoencoders) are effective at detecting side tasks when models are instructed to perform them, but they fail to detect such behaviors when they are internalized through fine-tuning. The study shows that NLAs produce more incriminating readouts in the In-Context Learning (ICL) modality compared to the SFT (Supervised Fine-Tuning) modality, where the behavior is internalized. The presence of instructions in the context window significantly influences the detection rate, as NLAs primarily detect verbalized instructions rather than covert computations.