Reproduce “PGuard: A Native, Streaming, and Plug-in Guardrail for 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
Junyu Lin, Mengjie Wu, Yonghao Dong, Yuchuan Fu, Jinfeng Li, Xiaodan Li, Xiaohan Yuan, Yicheng Liu and 6 more
PGuard monitors streaming answers by reading a model’s internal representations and sending category-specific risk scores to an application, which chooses whether to act. It attaches to a frozen backbone without changing its inference path. On Qwen3.5-9B text-query benchmarks, EMA and GRU each achieve 0.836 average F1 across six benchmarks, close to a different-backbone baseline; this does not show universal detection or prevention.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.