Signal CABAL: Multi-Agent Simulacra for Tracing the Effects of Collusive Bidding in Peer Review

Summary

This paper addresses the problem of reviewers colluding on bids, an issue raised during the AAAI-27 review cycle. The authors note a gap in prior work, which treated bidding, assignment, and review manipulation as separate stages, leaving the overall effect of collusion across the process unclear. To fill this gap they propose CABAL, a multi-agent simulation that holds the conference setting fixed while assigning LLM-based reviewer agents either honest or collusive policies. They also developed an affinity-guided collusive bidding strategy that uses mutual reviewer-paper affinity to form collusion rings and pick target papers. In controlled experiments, collusive bidding more than doubled the chance of capturing a target paper, and colluding reviewers scored those papers roughly two points higher than honest co-reviewers did, while the effect on the conference as a whole remained comparatively small. Existing detectors were confused by ordinary, non-collusive affinity and could only identify collusion precisely within a narrow scope.