Dude Uses Dual-Detection Multi-Agent System to Find Paper-Code Discrepancies

Summary

The paper introduces Dude, described as the first dual-detection multi-agent system for detecting discrepancies between research papers and their associated code. The work addresses a growing review problem: research submissions are scaling beyond the capacity of manual checking, while single-agent LLM approaches are limited by context capacity and tend to inspect discrepancies from only one direction. The authors also identify a granularity mismatch between natural-language papers and code as a source of over-interpretation and over-reporting in multi-agent designs, which can increase false positives. Dude therefore uses granularity-aligned negotiation between agents and a two-stage salience-filtering mechanism to suppress discrepancies that should not be reported. Experiments on real-world paper-code discrepancy datasets show improvements in both recall and precision of up to 22.8% over baseline methods, with F1 improving by up to 18.7%.