CABAL Simulates Peer-Review Collusion Rings and More Than Doubles Target-Paper Capture

arXiv 2609.05227·medium signalXBlueskyLinkedInCopy link

arXiv 2609.05227 responds to reports of coordinated reviewer bidding during the AAAI-27 review cycle by building an end-to-end multi-agent simulacra framework that holds the conference environment fixed and configures LLM-driven reviewer agents with honest or collusive policies. Its affinity-guided bidding strategy uses mutual reviewer-paper affinities to construct rings and pick targets, producing expertise-consistent rather than arbitrary attacks. Collusive bidding more than doubles target-paper capture and assigned colluders score targets about two points higher than honest co-reviewers, while evaluated bid-phase detectors give only limited evidence because native positive-bid graphs are confounded by benign affinity.