KG2Code Enhances LLM Question Answering with Knowledge Graph Code.

Integrating knowledge graphs (KGs) with large language models (LLMs) has been a focus of recent research to boost performance on knowledge-intensive tasks, particularly knowledge graph question answering (KGQA). While existing methods like RAG-based, agent-based, and SPARQL-based approaches have seen success, they often suffer from issues such as structural information loss, unfaithful reasoning, and limited flexibility. To overcome these challenges, researchers propose KG2Code, a new approach that transforms knowledge graphs into a code-based representation. This method effectively preserves the structural semantics of KGs while naturally aligning with the code-aware pretraining of modern LLMs. Building on KG2Code, the KG2Code-QA framework formulates KGQA as a code generation task. This formulation allows for the generation of verifiable reasoning traces and executable code, which significantly helps in reducing LLM hallucinations. Furthermore, an automated pipeline was developed to create a large, high-quality code corpus, enabling effective training of open-source LLMs on KG2Code-QA. After training, these LLMs can perform KGQA in zero-shot scenarios. Extensive experiments confirm that KG2Code-QA substantially outperforms existing KG-enhanced LLM methods for KGQA and demonstrates strong generalization to previously unseen knowledge graphs.