Go to ACL ARR 2026 January homepagePD$^3$: A Framework for Project Duplication Detection via Adapted Multi-Agent Debate
Keywords: Project Duplication Detection, Multi-Agent Debate, Large Language Model
Abstract: Project duplication detection is critical for project quality assessment by preventing investing in newly proposed projects whose topics are already studied. Existing methods rely on word- and sentence-level comparison or solely apply large language models, lacking valuable insights generation and in-depth comprehension of core content and detection criteria. To address this, we propose **PD$^3$**, a framework for **P**roject **D**uplication **D**etection via adapted multi-agent **D**ebate. Inspired by real-world expert discussion and tournament formats, PD$^3$ employs a fair round-robin competition format in multi-agent debate to retrieve relevant projects and generates both qualitative and quantitative feedback for greater practicality. Over 800 real-world power projects spanning 22 specialized fields are used for evaluation, demonstrating 8.37% and 8.00% improvements in precision and accuracy in two key tasks. Furthermore, we develop and deploy ***Review Dingdang***, an online platform assisting power experts, which has already saved 13.44 million USD across 442 newly proposed projects. Codes are available in [this repository](https://anonymous.4open.science/r/Anonymous-Repo-F1DB/).
Paper Type: Long
Research Area: NLP Applications
Research Area Keywords: NLP for social good, applications, user-centered design, LLM/AI agents,
Contribution Types: NLP engineering experiment, Reproduction study, Publicly available software and/or pre-trained models
Languages Studied: Chinese, English
Submission Number: 1258
Loading