EVENT DOSSIER
MABPD is a new approach for bias detection and identification in multi-agent systems without supervision. This approach consists of three specialized large language model agents, which analyze text content through a structured deliberation debate (SAD) protocol. In the BABE benchmark, this approach achieved a macro F1 score of 83.4%. Its core mechanisms include applying domain-driven asymmetric burden of proof and a role-weighted voting mechanism. A ablation experiment showed that removing the debate module resulted in a decrease of up to 10.6 points in the F1 score, proving that structured deliberation is crucial for performance improvement. In zero-sample evaluations on the SemEval 2019 HyperPartisan corpus, the accuracy was 75.0%, a difference of 7.2 percentage points from the most advanced supervised methods. The authors have made the entire processing pipeline and evaluation code available publicly.
arXiv:2609.04841v1 MABPD 发布基于结构化辩论的多智能体偏见检测管道
MABPD 提出一种无需监督训练的多智能体偏见探测与检测方案，在 BABE 基准测试中达到 83.4% 的宏观 F1 分数。该方案由三个专用 LLM 智能体组成，通过结构化论证辩论（SAD）协议分析文章，实施领域驱动的不对称举证责任及角色加权投票机制。消融实验证实，移除辩论模块会使 F1 下降最高达 10.6 分，表明结构化 deliberation 而非单纯并行提升性能。在 SemEval 2019 HyperPartisan 语料库上的零样本评估中，其准确率为 75.0%，与监督最先进方法差距为 7.2 个百分点。作者已公开完整管道及评估代码。