CABAL: A Multi-Agent Simulation Framework for Tracing the Impact of Collusive Bidding in Peer Review

The researchers propose an end-to-end multi-agent simulation framework, CABAL, which configures LLM reviewer agents with honest or collusive strategies in a fixed conference setting to study the impact of collusive bidding on the integrity of review assignment. The research background is reports from the AAAI-27 review cycle about reviewers coordinating bids to gain assignment advantages for each other. The framework proposes an affinity-guided collusive bidding strategy that uses the mutual affinity between reviewers and papers to build collusion rings and select target papers, making the attack consistent with professional direction rather than randomly targeted. Controlled experiments show that collusive bidding more than doubles the proportion of target papers captured by colluders, and that assigned colluders give target papers scores about two points higher than honest co-reviewers, while the conference-wide impact is relatively limited. The evaluated bidding-stage detectors can only provide limited evidence of collusion; in a stress test of a fixed triplet detector, the positive bidding graph is confounded by benign affinity, and only the diagnostic view retaining Very-High bids achieves precise but low-coverage local recovery.