Reproduce “ThinkingGuard: Decoding Implicit Hazards via Step-by-Step Risk Attribution in Multimodal Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruochen Zhang, Yao Huang, Yitong Sun, Jiahe Xie, Jin Yan, Jifan Ma, Yuanfang Guo, Xingxing Wei
ThinkingGuard detects multimodal safety risks that depend on interactions between text requests and visual context. It introduces TriggerBench, pairing safe and unsafe examples that change the environmental trigger while keeping the query and key entity fixed. The model analyzes intent, entities, attributes, and relationships before judging safety. It reaches 0.764 average F1-Unsafe and 0.866 recall on TriggerBench, not evidence of deployment performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.