Vision-Language Models Can Pass Hidden Preferences Through Unrelated Numbers

Introduction

Can a model learn a teacher’s hidden preference even when the teacher only produces numbers for an unrelated visual task? A study on subliminal learning in vision-language models investigates exactly this possibility. Subliminal learning describes a form of behavioral transfer in which a teacher’s latent preference reaches a student through data that does not explicitly mention, label, or reward that preference.

Key findings

• The researchers fine-tuned teacher vision-language models to prefer a specific animal or a particular facial attribute. A control teacher was created from the same base model.
• Both teachers then handled a visual task involving geometric shapes and generated numerical outputs. The task itself was unrelated to animals and facial characteristics.
• Two student models, initialized from the same base VLM, were trained on the same images but used outputs from either the biased teacher or the control teacher.
• Across VLMs ranging from 2B to 7B parameters, students trained on biased-teacher outputs showed the corresponding preference. In the animal setting, the difference from the control group reached roughly 10 percentage points.
• The transfer was not limited to final choices. Even when the models selected the same image, the biased student’s explanation shared more words and phrases with the biased teacher’s explanation.

Why it matters

The result broadens subliminal learning from language models to multimodal systems. It suggests that a model’s influence can be embedded in apparently ordinary intermediate outputs, including numbers generated for a task with no obvious connection to the teacher’s learned preference. In knowledge distillation, such outputs should therefore not automatically be treated as behaviorally neutral.

The strength of the effect varied with the target preference, the model, and the way numbers were generated. It could also increase with more training data. These observations indicate that the phenomenon is not a single, uniform mechanism; its reliability and boundary conditions still require investigation.

For practitioners, the findings raise a practical evaluation issue. Checking whether a student correctly performs the nominal numerical task may miss changes in its visual preferences or explanation style. Multimodal training pipelines may need cross-task behavior tests, preference audits, and closer examination of teacher-generated supervision.

The study does not imply that every distillation process will produce a visible bias. Its broader warning is more specific: data that is unrelated at the task level may still carry information about a teacher’s internal behavioral tendencies. Those tendencies can survive training and reappear when the student makes new visual judgments.

Source: Hugging Face Daily Papers