Research Papers 论文研究 6h ago Updated 1h ago 更新于 1小时前 45

Dude: A Dual-Detection Multi-Agent System for Paper-Code Discrepancy Detection Dude:一种用于论文-代码差异检测的双检测多智能体系统

Dude is the first dual-detection multi-agent system designed to identify discrepancies between academic papers and their associated code implementations The system addresses granularity asymmetry between natural language papers and code, which causes over-interpretation and over-reporting of false positives in multi-agent designs A granularity-aligned negotiation mechanism and a two-stage salience-filtering mechanism are introduced to reduce false discrepancy reports Experimental results show up 提出Dude,首个双检测多Agent系统用于论文-代码差异检测,解决单Agent方法上下文容量有限和召回率低的问题 发现论文语言与代码语言的粒度不对称会导致多Agent系统出现过解释和过度报告,进而增加误报 设计粒度对齐协商机制与两阶段显著性过滤机制,有效抑制Agent的虚假差异报告 在真实论文-代码差异数据集上,Dude的召回率和精确率最高提升22.8%,F1分数最高提升18.7%

60
Hot 热度
72
Quality 质量
62
Impact 影响力

Analysis 深度分析

TL;DR

  • Dude is the first dual-detection multi-agent system designed to identify discrepancies between academic papers and their associated code implementations
  • The system addresses granularity asymmetry between natural language papers and code, which causes over-interpretation and over-reporting of false positives in multi-agent designs
  • A granularity-aligned negotiation mechanism and a two-stage salience-filtering mechanism are introduced to reduce false discrepancy reports
  • Experimental results show up to 22.8% improvement in recall and precision, and up to 18.7% increase in F1 score compared to baseline methods
  • The work is motivated by the growing volume of research submissions that exceed manual review capacity, making LLM-empowered automated detection increasingly necessary

Why It Matters

As AI research submissions continue to scale, manually verifying that published code matches paper claims has become infeasible, creating a critical need for automated discrepancy detection. This work directly addresses a gap in AI research integrity by introducing a multi-agent architecture that improves both recall and precision, which is essential for trustworthy automated peer review and reproducibility verification.

Technical Details

  • Dual-Detection Multi-Agent Architecture: Dude employs two complementary detection agents that cross-validate findings, addressing the one-sided detection limitations of single-agent LLM paradigms
  • Granularity-Alignment Negotiation: The system introduces a negotiation mechanism that aligns the different granularity levels between paper language (high-level descriptions) and code (low-level implementations), reducing over-interpretation errors
  • Two-Stage Salience-Filtering Mechanism: A filtering pipeline that first identifies potentially discrepant claims and then re-evaluates their salience to suppress false positives before final reporting
  • Benchmark Performance: Evaluated on real-world paper-code discrepancy datasets, achieving up to 22.8% improvement in both recall and precision, and up to 18.7% F1 score gain over baselines
  • Context Capacity Consideration: The multi-agent design accounts for limited context windows in LLMs, distributing the detection task across agents to overcome single-agent context constraints

Industry Insight

  • The increasing volume of AI research submissions makes automated verification tools like Dude essential for maintaining research integrity and reproducibility standards in the community
  • Multi-agent LLM systems that address granularity mismatches between modalities (text vs. code) represent a generalizable pattern applicable to other cross-modal verification tasks beyond paper-code comparison
  • The demonstrated improvement in both precision and recall suggests that dual-detection architectures could become a standard pattern for automated peer review and code-paper validation pipelines in academic publishing

TL;DR

  • 提出Dude,首个双检测多Agent系统用于论文-代码差异检测,解决单Agent方法上下文容量有限和召回率低的问题
  • 发现论文语言与代码语言的粒度不对称会导致多Agent系统出现过解释和过度报告,进而增加误报
  • 设计粒度对齐协商机制与两阶段显著性过滤机制,有效抑制Agent的虚假差异报告
  • 在真实论文-代码差异数据集上,Dude的召回率和精确率最高提升22.8%,F1分数最高提升18.7%

为什么值得看

随着研究投稿量激增,人工审核已无法覆盖,自动化的论文-代码一致性检测成为AI辅助科研的重要需求。Dude通过多Agent协作与双重检测机制,为学术诚信审查和可重复性验证提供了可扩展的技术路径。

技术解析

  • 双检测架构:系统同时从论文描述和代码实现两个维度进行差异检测,弥补单Agent单向检测的不足
  • 粒度对齐协商机制:针对论文语言(高层语义)与代码语言(底层实现)的粒度不对称问题,设计跨粒度对齐策略,减少因抽象层级不匹配导致的误判
  • 两阶段显著性过滤:第一阶段粗筛候选差异,第二阶段基于显著性评分过滤低置信度报告,有效降低过度报告(over-reporting)问题
  • 实验验证:在真实论文-代码差异数据集上对比基线方法,Dude在召回率、精确率和F1分数上均取得显著提升

行业启示

  • 多Agent系统在复杂语义对齐任务中展现出超越单Agent的潜力,未来可推广至代码审查、学术诚信检测等场景
  • 粒度对齐与显著性过滤是抑制多Agent系统误报的关键设计原则,对构建高可靠性Agent协作系统具有通用参考价值
  • 随着AI生成代码和论文的普及,自动化一致性验证工具将成为科研基础设施的重要组成部分

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Agent Agent Research 科学研究 Evaluation 评测 Code Generation 代码生成