Zhang, Zetong (The University of Queensland), Kim, Jiwon (The University of Queensland), He, Dan (University of Queensland), Yildirimoglu, Mehmet (University of Queensland)

Reflective LLM Prompt Optimisation for Interpreting GNN Predictions in Traffic Forecasting

Scheduled for presentation during the Invited Session "S18b-Innovative Applications of LLM in Multimodal Transportation Systems" (TH-EA-T18), Thursday, November 20, 2025, 14:30−14:50, Southport 3

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Graph Neural Networks (GNNs) are effective for traffic forecasting, but their predictions are difficult to interpret. While post-hoc explanation tools like GNNExplainer generate useful importance scores across graph nodes and edges, the raw outputs are hard to interpret without additional context. Large Language Models (LLMs) offer a way to translate these scores into natural language, but often fluent yet unfaithful explanations-even when the input scores are invalid. We propose a two-phase, multi-agent reflective framework to address these limitations. Phase 1 optimises a prompt that enables LLMs to assess the validity of GNN importance scores using synthetic labelled data. Phase 2 incorporates this optimised prompt and focuses on refining the explanation component of the prompt to produce more faithful narratives grounded in multimodal traffic context. Experiments show that our framework improves the LLM's ability to validate input scores before generating explanations as well as the ability to generate more trustworthy, context-aware explanations grounded in multimodal evidence. The framework provides a structured path toward more adaptable and dependable LLM-based interpretation of complex machine learning models in critical traffic prediction applications.