Subliminal Learning in Vision-Language Models
Students trained on a biased teacher’s number outputs chose its preferred animal up to roughly 10 percentage points more often than controls.
The teachers generated numbers for an unrelated task involving images of geometric shapes. Across vision-language models with 2B–7B parameters, both animal and facial-attribute preferences transferred to students trained on those outputs. Transfer sometimes strengthened with more training data and varied by model, target preference and number-generation method. Students also echoed more of the biased teacher’s wording in explanations, even when biased and control models chose the same image.
HF Daily Papers' note
The teachers generated numbers for an unrelated task involving images of geometric shapes. Across vision-language models with 2B–7B parameters, both animal and facial-attribute preferences transferred to students trained on those outputs. Transfer sometimes strengthened with more training data and varied by model, target preference and number-generation method. Students also echoed more of the biased teacher’s wording in explanations, even when biased and control models chose the same image.
HF Daily Papers' note
score 5