DataDecide OLMo

Subliminal learning persists in LLMs across training variations

Researchers have investigated "subliminal learning" in large language models, where models trained on data generated by other models can inherit unintended traits. The study explored both hard label distillation, using …