Subliminal Learning in Vision-Language Models
Paper · https://arxiv.org/abs/2611.08538
arXiv:2611.08538v1 Announce Type: new Abstract: Subliminal learning is a phenomenon in large language models (LLMs) where a teacher passes a hidden preference to a student through training data unrelated to that preference. We study whether subliminal learning also occurs in vision-language models (VLMs). We finetune a teacher VLM to prefer a specific animal or facial attribute and use a control teacher from the same base model. Both teachers generate numbers for an unrelated visual task involving geometric shapes. Two students initialized from the same base VLM are finetuned on the same images, using numbers generated by either the biased or control teacher. Across VLMs ranging from 2B to 7B parameters, we find that both animal and facial preferences can transfer through this unrelated task. Students trained on biased-teacher outputs select the target more often than control students, with a difference of up to $\sim$10 percentage points in the animal preference task. Transfer can become stronger with more training data and varies with the target preference, model, and how the numbers are generated. Beyond preferences, we show that the biased student's explanations use more of the same words and phrases as the biased teacher's explanations than those of the control student, even when all three models select the same image. These results matter for training VLMs and show that training on unrelated data may change both a VLM's image preferences and how it explains its choices, and that transfer can become stronger with more training samples.
First seen by Oriel on .
AI summary
The paper finds that subliminal learning occurs in vision-language models, transferring hidden animal and facial preferences through numerical outputs from an unrelated geometric-shape task.
- Researchers finetuned biased and control teachers from the same base VLM, then trained identically initialized students on numbers generated for identical images.
- The effect appeared across VLMs from 2B to 7B parameters, reaching roughly a 10-percentage-point difference in target animal selection.
- Transfer strength increased with more training data in some cases and varied by target preference, model, and number-generation method.
- Biased students also echoed more words and phrases from biased-teacher explanations, even when biased, control, and teacher models selected the same image.
#vision-language models #subliminal learning #model finetuning #model preferences
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| arXiv AI Papers | oai:arXiv.org:2611.08538v1 | https://arxiv.org/abs/2611.08538 Read discussion |