LLMs transmit behavioral traits through hidden signals

This discovery undermines the safety of synthetic data distillation and model fine-tuning by demonstrating that a teacher model's biases can "infect" a student through unrelated data. The "Owl Experiment" provides empirical proof that arbitrary traits leak through parameter-level signals, making synthetic data a potential vector for "hidden contagion" of misaligned behaviors. Theoretical results confirm that gradient descent on teacher-generated data moves students toward the teacher's parameter space, implying that AI safety must evolve beyond behavioral evaluation to include rigorous audits of training data origins.