← All projects

When Graph Structure Hurts

Lightweight path ranking for dense KG-RAG · ICMLC 2026

Knowledge GraphsRAGPyTorchGNN

Overview

Research accepted at ICMLC 2026 (presentation, Feb 2026). We study retrieval over dense academic citation graphs and ask: when does graph structure actually help KG-RAG?

Key result

A lightweight MLP-based path scorer achieves 93.9% AUC, outperforming GCN and GAT models while using 13× fewer parameters. In dense graphs, heavy message-passing architectures can amplify noise rather than signal.

Method

  • KG-augmented RAG pipeline over citation graphs
  • Candidate path generation, then learned ranking of paths feeding the retriever
  • Systematic comparison: MLP scorer vs. GCN vs. GAT under matched budgets

Why it matters

Most KG-RAG work assumes more graph machinery = better retrieval. This shows the opposite can hold in dense regimes, useful for anyone deploying KG-RAG where inference cost matters.