Dude: A Dual-Detection Multi-Agent System for Paper-Code Discrepancy Detection

Forum topic · 小凯 · 2026-09-06

Summary

Dude is a dual-detection multi-agent system designed to detect discrepancies between research papers and their accompanying code, addressing the scalability limits of manual review as research submissions grow. The authors identify two core weaknesses in existing single-agent LLM approaches: limited context capacity and one-sided discrepancy detection, both of which hurt recall. They further show that granularity asymmetry between paper language and code language causes over-interpretation and over-reporting in multi-agent designs, increasing false positives. Dude mitigates this with a granularity-aligned negotiation mechanism and a two-stage salience-filtering process. Experiments on a real-world paper-code discrepancy dataset show that Dude improves recall and precision by up to 22.8% and boosts F1 score by up to 18.7% over baseline methods. The paper is available on arXiv (2509.00007).

Overview

Field: AI/ML Authors: Weijie Liu, Running Zhao, Wenhao Yuan Published: 2026-09-06 arXiv: 2509.00007

Abstract

LLM-empowered paper-code discrepancy detection has received growing concern since the scaling of research submissions exceeds the manual review capability. However, the limited context capacity and one-sided discrepancy detection of existing single-agent LLM paradigms lead to an inferior recall performance in detecting discrepancies. In this paper, we propose Dude, the first Dual-Detection Multi-Agent System for paper-code discrepancy detection.

Key Contributions

Dual-detection design: Dude uses multiple agents to analyze both the paper and the code, overcoming the limited context capacity and one-sided detection of single-agent LLM paradigms.
Granularity-aligned negotiation: The authors observe that the granularity asymmetry between paper-language and code-language introduces over-interpretation and over-reporting challenges in multi-agent designs, increasing false positives. A negotiation mechanism aligns these granularities.
Two-stage salience filtering: This mechanism further prevents agents from misreporting discrepancies.

Results

Experiments on a real-world paper-code discrepancy dataset show that Dude significantly improves recall and precision by up to 22.8%, with an F1 score improvement of up to 18.7% over baseline methods.

--- *Auto-collected on 2026-09-06*