Subliminal learning transfers preferences in vision-language models

Researchers have investigated the phenomenon of subliminal learning in vision-language models (VLMs), where a teacher model imparts hidden preferences to a student model through unrelated training data. Experiments involving VLMs of varying sizes (2B to 7B parameters) demonstrated that biases in animal or facial attributes could be transferred from a teacher to a student model, even when the training task focused on geometric shapes. This transfer resulted in the student model selecting the target attribute more frequently, with differences up to approximately 10 percentage points in some cases. Furthermore, the study found that the biased student's explanations mirrored the teacher's phrasing more closely than a control student's, indicating changes in both preference and explanatory style. AI

IMPACT Reveals potential for hidden biases to transfer in VLMs, impacting training data curation and model interpretability.

RANK_REASON The cluster contains a research paper detailing a new finding about model behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• ScienceCast
• subliminal learning
• vision-language model
• Vision--Language Models

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →