CABAL: Multi-Agent Simulacra for Tracing the Effects of Collusive Bidding in Peer Review
A paper titled 'CABAL: Multi-Agent Simulacra for Tracing the Effects of Collusive Bidding in Peer Review' was published on arXiv (cs.AI) on 2026-09-04. The paper introduces CABAL, an end-to-end multi-agent simulacra framework for studying reviewer assignment integrity. It develops an affinity-guided collusive bidding strategy. Controlled experiments show collusive bidding more than doubles target-paper capture and assigned colluders score target papers about two points higher than honest co-reviewers, while conference-wide effects remain comparatively modest.
The paper presents CABAL, a multi-agent simulation framework to study collusive bidding in peer review. It uses LLM-driven reviewer agents with honest or collusive policies, holding the conference environment fixed. The affinity-guided collusive bidding strategy constructs collusion rings and selects target papers based on mutual reviewer-paper affinities. Experiments show collusive bidding significantly increases target-paper capture and inflates scores for target papers, but overall conference-wide impact is modest.
CABAL uses LLM-driven reviewer agents configured with honest or collusive policies within a fixed conference environment. The affinity-guided collusive bidding strategy leverages mutual reviewer-paper affinities to form collusion rings and select target papers, producing expertise-consistent attacks. The framework enables counterfactual analysis of reviewer assignment integrity, which is difficult with real-world data due to unobservable collusive intent.
The research addresses integrity risks in AI conference peer review, particularly collusive bidding. It suggests that while collusion can significantly affect targeted papers, the overall conference-wide impact is limited. This may inform the design of detection and mitigation strategies for academic conferences and other peer review systems.
The framework could be used by conference organizers and publishers to test and improve the integrity of their review processes. It may also be relevant to platforms that rely on peer review or expert evaluation, such as grant agencies or journal publishers, to assess vulnerabilities to collusion.
The paper mentions evaluated bid-phase detectors, indicating ongoing work on detection methods. Future research may extend the framework to other stages of the review process or apply it to different conference settings. The approach could also be adapted to study other forms of strategic manipulation in peer review.