This paper addresses the challenge of applying Large Language Models (LLMs) to spatial-temporal forecasting, a task that is crucial in various fields such as traffic, weather, and epidemic modeling. The authors introduce STG-LLM, an innovative approach that combines a spatial-temporal graph tokenizer (STG-Tokenizer) and a spatial-temporal graph adapter (STG-Adapter) to enable LLMs to understand and predict spatial-temporal data effectively. The STG-Tokenizer transforms complex graph data into concise tokens that capture both spatial and temporal relationships, while the STG-Adapter, consisting of linear encoding and decoding layers, bridges the gap between tokenized data and LLM comprehension. By fine-tuning only a small set of parameters, the adapter can effectively grasp the semantics of the tokens generated by the tokenizer, preserving the original natural language understanding capabilities of the LLMs. Experiments on diverse spatial-temporal benchmark datasets demonstrate that STG-LLM successfully unlocks the potential of LLMs for spatial-temporal forecasting, achieving competitive performance comparable to state-of-the-art methods. The approach is particularly effective in handling data sparsity and generalizing to new datasets with limited training data. Additionally, the paper explores the effectiveness of prompts and conducts an ablation study to validate the key components of STG-LLM, showing that the proposed method is robust and efficient.This paper addresses the challenge of applying Large Language Models (LLMs) to spatial-temporal forecasting, a task that is crucial in various fields such as traffic, weather, and epidemic modeling. The authors introduce STG-LLM, an innovative approach that combines a spatial-temporal graph tokenizer (STG-Tokenizer) and a spatial-temporal graph adapter (STG-Adapter) to enable LLMs to understand and predict spatial-temporal data effectively. The STG-Tokenizer transforms complex graph data into concise tokens that capture both spatial and temporal relationships, while the STG-Adapter, consisting of linear encoding and decoding layers, bridges the gap between tokenized data and LLM comprehension. By fine-tuning only a small set of parameters, the adapter can effectively grasp the semantics of the tokens generated by the tokenizer, preserving the original natural language understanding capabilities of the LLMs. Experiments on diverse spatial-temporal benchmark datasets demonstrate that STG-LLM successfully unlocks the potential of LLMs for spatial-temporal forecasting, achieving competitive performance comparable to state-of-the-art methods. The approach is particularly effective in handling data sparsity and generalizing to new datasets with limited training data. Additionally, the paper explores the effectiveness of prompts and conducts an ablation study to validate the key components of STG-LLM, showing that the proposed method is robust and efficient.