The rapid growth of research submissions has made paper‑code discrepancy detection powered by large language models (LLMs) increasingly important. Existing single‑agent LLM approaches suffer from limited context windows and one‑sided checking, which harms recall.\ \ We introduce Dude, the first dual‑detection multi‑agent system for this task. We observe that the granularity mismatch between natural‑language papers and programming‑language code leads to over‑interpretation and over‑reporting, inflating false positives. Dude tackles these issues with two core techniques:\
- Granularity‑aligned negotiation – the agents exchange granularity cues so that they compare paper statements and code snippets at the same abstraction level.\
- Two‑stage salience filtering – the first stage discards low‑semantic‑similarity fragments; the second stage applies structured prompts to re‑validate the remaining candidates, keeping only high‑confidence discrepancies.\ On real‑world paper‑code discrepancy benchmarks, Dude improves recall and precision by up to 22.8% and raises the F1 score by 18.7% over strong baselines.\ \ Review