Subliminal Learning in Vision-Language Models

Subliminal Learning in Vision-Language Models. By Arman Hatami and two others, from Johns Hopkins University. Published on the fourth of November.

Subliminal learning is a phenomenon described in large language models. A teacher passes a hidden preference to a student through training data that has nothing to do with that preference. The question here is whether the same thing happens in vision-language models, and the answer reported is yes. Preferences for a particular animal, and for a particular facial attribute, can pass from a teacher VLM to a student through an unrelated visual task.

The setup turns on a careful control. One teacher VLM is finetuned to prefer a specific animal or a specific facial attribute. A control teacher comes from the same base model. Both teachers are then given the same unrelated job: generating numbers for a visual task involving geometric shapes. Nothing about animals or faces is part of it. Two students, each initialized from the same base VLM, are finetuned on the same images. The only difference between them is whose numbers they learn from, the biased teacher's or the control teacher's.

Across models ranging from two billion to seven billion parameters, students trained on the biased teacher's outputs select the target more often than control students do. In the animal preference task the difference reaches about ten percentage points. The transfer shows up in more than the choices. When the biased student explains itself, its explanations use more of the same words and phrases as the biased teacher's explanations than the control student's do. That holds even in cases where all three models select the same image, so the trace is in how the choice is explained, not only in what is chosen.

The effect is not uniform. Transfer varies with the target preference, with the model, and with how the numbers are generated, and it can become stronger with more training data. The models tested run from two billion to seven billion parameters. The preferences are animals and facial attributes, and the carrier task is number generation over images of geometric shapes. The figure of about ten points is the largest difference reported, and it belongs to the animal task specifically rather than to every condition.

For anyone training VLMs, the implication the authors draw is direct. Training on data that looks unrelated may still change which images a model prefers and how it explains its choices, and more training samples can make that change stronger. Data being unrelated in content, on this evidence, does not guarantee that nothing else rides along with it.