Propagation structures provide vital clues for fake‑news detection, yet most existing methods rely on supervised GNNs that demand large labeled corpora and suffer from limited generalization. Large language models (LLMs) possess strong reasoning abilities, but feeding raw propagation graphs directly to them creates a severe modality mismatch and information overload, rendering structure‑aware reasoning unreliable in zero‑shot or few‑shot scenarios.
To bridge this gap, we introduce MAGER (Multi‑Agent Genetic Evolution for Reasoning), a multi‑agent genetic evolution framework that automatically discovers meta‑paths optimized for LLM reasoning. By compressing complex propagation graphs into informative subgraphs, the evolved meta‑paths alleviate both information overload and modality mismatch, enabling frozen LLMs to perform structure‑aware veracity reasoning.
We further propose a graph in‑context learning strategy that retrieves semantically and structurally similar demonstrations as in‑context examples, strengthening classification and reasoning robustness.
Extensive experiments on real‑world datasets demonstrate that MAGER substantially improves frozen LLMs as standalone fake‑news detectors in data‑efficient settings. The code is released at https://github.com/SenticNet/MAGER.
Review