Reproduce “Judging Before Propagation: Task-Graph Contextual Integrity for Multi-Agent Privacy” 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
Hengyu Wu, Blaise Delattre, Yang Cao
TG-CI extends contextual integrity checks to multi-agent workflows by evaluating an item against recipients’ disclosure histories and its planned downstream task path. Full-Scope blocks an item if any planned transfer is inadmissible, while Auto-Region removes it at the first inadmissible edge. On synthetic MAPLE-Bench, unauthorized-disclosure events fall from 10.47 to 1.49 and 0.62; these results do not establish real-world effectiveness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.