发表机构
Rutgers University, New Brunswick; University of Cambridge; Tsinghua University; Harvard University; Case Western Reserve University; University of California San Diego; Carnegie Mellon University（罗格斯大学新布朗斯维克分校; 剑桥大学; 清华大学; 哈佛大学; 凯斯西储大学; 加利福尼亚大学圣迭戈分校; 卡内基梅隆大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
提出条件渐进剪枝（CPP）框架，通过轻量级剪枝充分利用多轮多智能体辩论，在严格成本限制下首次全面超越单智能体与一致性基线，并在多个基准上优于现有MAD方法。
AI 中文摘要
基于大型语言模型（LLM）的多智能体辩论（MAD）是最有效的测试时扩展技术之一。通过多轮通信，智能体在知识和推理方面相互补充，解决单个成员无法解决的问题。然而，现有的MAD框架在相同的严格成本限制下未能击败强大的单智能体和基于一致性的基线，这动摇了MAD领域的基础。我们提出条件渐进剪枝（CPP），一种轻量级剪枝框架，充分利用多轮MAD。CPP在多个主导基准上优于所有现有MAD框架，并且是首个全面超越一致性方法的框架。我们的代码和详细的智能体交互记录将很快发布。
英文摘要
Large Language Model (LLM) based Multi-Agent Debate (MAD) is one of the most effective test time scaling techniques. Through multi-round communication, agents complement each other in knowledge and reasoning and solve tasks that no single member can solve. However, existing MAD frameworks fail to beat strong Single Agent and Consistency-based baselines under the same strict cost limit, which shakes the foundation of the MAD field. We propose Conditional Progressive Pruning (CPP), a lightweight pruning framework that fully exploits multi-round MAD. CPP outperforms all existing MAD frameworks on multiple dominated benchmarks. It is also the first to fully outperform consistency methods. Our code, detailed agent interaction records will be released soon.
