Reproduce “Multi-Agent Collusion and the J-Space” 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
Marco Molinari, Gabriele Bettineschi
J-PACT compares shared concepts between agent pairs and their other agents’ intermediate activations to detect collusion without labels. It requires same-model agents and assumes honest agents have comparable roles. Its four-model mean run-level AUROC is 0.921 on Core versus 0.993 for supervised probes, and 0.991 on Stego; on role-structured Transfer, with one observation per agent, it is 0.519, near chance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.