Does AI Reviewer See the Full Picture? Attacking and Defending Multimodal Peer Review

In three linesPaperGuard, a multimodal benchmark, evaluates LLMs and MLLMs vulnerability to adversarial attacks in scientific peer review. Researchers test prompt injections and perturbations (GCG, PGD) on text and figures, proposing a defense using chunk-based embedding search to localize harmful instructions.

Summary generated by Claude — human-verified