视觉语言模型中的潜意识学习
原标题：Subliminal Learning in Vision-Language Models
AI 摘要
研究者在视觉语言模型（VLM）中验证了「潜意识学习」现象：先微调一个教师 VLM 使其偏好特定动物或面部属性，再让该教师为与偏好无关的几何图形计数任务生成数字，并用这些数字微调学生 VLM。结果显示，2B 至 7B 参数的 VLM 均能将隐藏偏好通过无关任务传递给学生，动物偏好任务中学生选择目标的比例最多高出约 10 个百分点，且传递效应随训练数据增加而增强。研究还发现，学生模型的解释用词会向教师靠拢，说明在无关数据上训练可能改变 VLM 的图像偏好及其解释方式。
正文节选
Subliminal Learning in Vision-Language Models 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