Multi-Agent Debate Framework Detects Duplicate Projects, Saving Millions in Investment

Project duplication—submitting similar or overlapping proposals for funding—wastes resources and creates inefficiencies in capital allocation. Identifying these duplicates before investment decisions are made is essential for maintaining portfolio quality, yet existing approaches often struggle to compare proposals fairly at scale.

The Challenge with Current Methods#

Conventional duplication detection systems typically rely on surface-level text matching or delegate the decision entirely to large language models without structured debate. These approaches have a critical flaw: they fail to consider the broader set of candidate projects when making comparisons, and they cannot guarantee fair evaluation across many proposals simultaneously.

The real-world problem requires what researchers call "many-to-many reference set selection"—the ability to sift through numerous proposals, understand their relationships, and determine which ones are genuinely duplicative while accounting for context and resource constraints.

How PD³ Works#

The new PD³ framework tackles duplication detection by combining two techniques: local multi-agent debate and global round-robin scheduling. Individual AI agents debate the merits and overlaps of pairs of projects, while a scheduling layer ensures each proposal gets equal exposure and comparison context. This hybrid approach generates both quantitative duplication scores and qualitative feedback explaining the overlap.

The framework's scheduler guarantees fairness by rotating projects through comparisons in a balanced way, preventing any single proposal from receiving skewed treatment due to ordering or context limitations.

Real-World Results#

When tested on more than 800 power generation projects, PD³ outperformed existing systems by 4.05% in selecting relevant reference sets and 9.77% in generating accurate duplication scores. The researchers deployed the system in Review Dingdang, a live platform for project assessment, which has already helped prevent wasteful spending by identifying duplicates across 442 new proposals—achieving documented savings of $13.44 million.

Why It Matters#

The success of PD³ demonstrates that structured AI reasoning—using coordinated agents and fair scheduling—can outperform both simple matching heuristics and unguided language models. Beyond power projects, this approach could improve duplication detection across funding mechanisms, grant systems, and innovation platforms where redundant proposals drain capital from potentially higher-impact work.