EVENT DOSSIER
CABAL proposes an end-to-end multi-agent simulation framework for studying the impact of collusive bidding on the integrity of review assignments during peer reviews. This framework constructs collusive bidding strategies based on affinity to form collusive loops and select target papers, by using LLM-driven reviewer agents that maintain a fixed conference environment and implement honest or collusive strategies. Controlled experiments show that collusive bidding doubles the capture rate of target papers; the assigned colluders score the target papers about two points higher than honest co-reviewers, while the overall impact on the conference is relatively limited. Additionally, there is limited evidence from existing bidding phase detectors: in stress tests with fixed triple-set detectors, native positive bidding graphs are susceptible to benign affinity confusion, and only the “very high” diagnostic view achieves accurate but low-coverage local recovery.
CABAL 提出一种端到端多智能体仿真框架，用于研究同行评审中合谋投标对审稿分配完整性的影响。该框架固定会议环境并配置具有诚实或合谋策略的 LLM 驱动审稿人代理，开发了基于亲和度的合谋投标策略以构建合谋环并选择目标论文。受控实验显示，合谋投标使目标论文捕获率翻倍，分配的合谋者给目标论文的评分比诚实共同审稿人高约两分，而会议整体影响相对有限。现有投标阶段检测器证据有限：在固定三元组检测器压力测试中，原生正投标图易受良性亲和度混淆，仅“极高”诊断视图能实现精确但覆盖率低的局部恢复。