Recent reports from the AAAI‑27 review cycle have highlighted the risk that reviewers may coordinate their bids to obtain reciprocal assignment advantages, posing a subtle threat to the integrity of peer‑review systems. Prior work typically treats bidding, reviewer assignment, and review manipulation as separate stages, leaving the lifecycle impact of collusive bidding unclear. Real‑world analysis is hampered by the difficulty of observing collusive intent and the lack of counterfactuals within the same conference.
To address this gap we introduce CABAL (\alg), an end‑to‑end multi‑agent simulacra framework. By fixing the conference environment and configuring LLM‑driven reviewer agents with either honest or collusive policies, CABAL enables systematic study of how bidding behavior propagates through assignment and review stages.
For the bidding stage we develop an affinity‑guided collusive strategy. The method builds collusion rings based on mutual reviewer‑paper affinities and selects target papers within those rings, ensuring that attacks are expertise‑consistent rather than arbitrarily targeted. Controlled experiments show that collusive bidding more than doubles the capture rate of target papers, and colluders assigned to those papers award scores about two points higher than honest co‑reviewers. While the impact on individual papers is pronounced, conference‑wide scoring shifts remain modest.
We also evaluate several bid‑phase detectors. A fixed‑triplet detector is confounded by benign affinity patterns in native positive‑bid graphs, whereas a "Very‑High‑only" diagnostic view achieves precise local recovery but with low coverage. Overall, existing detection mechanisms provide limited evidence of collusion.
Review