Announcement_13

[NeurIPS2025] We find patches from harmful content, enabling them to bypass data moderation and generate dangerous responses when encountering the full image or related text. VLMs Can Aggregate Scattered Training Patches is accepted by NeurIPS2025.