Researchers introduced subliminal learning as a phenomenon where a teacher model passes undetectable traits to a student model via a dataset. They presented SALVE, a method using text optimization to detect and verbalize these hidden traits. SALVE recovers legible prompts that identify the teacher’s trait in standard settings. It also works in mixed data, biased teacher models, and real preference data subsets. The method can reveal traits even when subliminal learning fails. Improving context distillation in student training can create subliminal learning effects. The paper deepens understanding of subliminal learning and offers a proactive detection method.