A teacher model that loves cats can generate a dataset of random-looking number sequences. Nothing in the text says “cats.” Fine-tune a student on it anyway, and the preference can still transfer. That is subliminal learning. It is also a data-poisoning problem, because the trait is not legible in the dataset. New Stanford work by Nathan Hu, Sanmi Koyejo, and Christopher Potts makes the hidden signal readable. They treat prompted subliminal learning as a special case of context distillation: in theory, the dataset identifies the teacher’s prompt. Then they recover that prompt with SALVE (Search-Aided Latent Verbalization): - Optimize a soft prompt so the original model predicts the dataset well - Ask the same model to verbalize the soft prompt - Use beam search to keep the wording fluent and faithful On the standard animal-preference setup, SALVE recovers prompts that name the trait in 18/20 runs. Common text-optimization baselines do not. In some cases it even recovers the trait from data where student fine-tuning fails to pick it up. They also detect the effect in mixed data, activation-steered teachers, and subsets of real preference data selected with Logit-Linear Selection (sycophancy / misalignment). The practical point is simple: you may be able to audit a dataset for hidden behaviors before it ever reaches fine-tuning. Paper: https://t.co/1mncXR5Fyv