When a language model is distilled from a teacher that shares its base parameters, the student can inherit the teacher’s behavioural traits—including misalignment—even if the training data are semantically unrelated to the trait (Cloud et al., 2026). Existing defenses all assume the trait is known in advance. We remove this assumption and show that the subliminally transferred trait can be recovered without supervision by looking only at the parameter updates. The key observation is a single asymmetry: the trait manifests consistently across diverse datasets generated by the teacher, while task‑specific and shared capabilities vary. Leveraging this, we design an estimator that returns a direction in weight space. Amplifying this direction reveals and interprets the trait; projecting it out suppresses the trait. On the subliminal‑learning benchmarks of Cloud et al. (2026) we recover the transmitted animal‑preference trait without any trait labels and suppress its expression with negligible impact on task performance.
Review: The approach offers a label‑free, generic defense by isolating a stable weight‑space direction, marking a significant step toward secure model distillation.