Traff-LLM: A Spatio-Temporal Knowledge-Guided Large Language Model for Traffic Flow Prediction

Traffic flow prediction is a critical task for intelligent transportation systems. However, existing deep learning models often suffer from poor interpretability and deployment challenges due to their architectural complexity. To address these limitations, we introduce Traff-LLM, a spatio-temporal knowledge-guided large language model (LLM) for accurate and interpretable traffic forecasting. Our framework first constructs a spatio-temporal knowledge graph to capture complex dependencies from traffic data. To manage the scale of this graph for LLM processing, we propose a novel Temporal Prize-Collecting Steiner Tree (T-PCST) algorithm that efficiently extracts compact, task-relevant subgraphs. Each subgraph and its associated external factors (e.g., weather and public events) are serialized into a unified natural-language prompt, which is then paired with the corresponding traffic flow to form the supervised fine-tuning dataset. The LLM is then fine-tuned on these prompt–response examples, enabling it to integrate multimodal signals and perform reasoning over the spatio-temporal graph. Experiments on two real-world benchmarks demonstrate that Traff-LLM significantly outperforms state-of-the-art methods, including specialized Graph Neural Networks and recent LLM-based approaches.