The article introduces a new method called GraphToken to encode structured data for use in large language models (LLMs). This approach is different from previous work which focused on limited domains like knowledge graph representation. The novelty lies in the general encoding of structured data for various reasoning tasks. The study shows that explicitly representing the graph structure allows significant improvements in graph reasoning tasks. The GraphToken method showed improvements up to 73% points on node, edge, and graph-level tasks from the GraphQA benchmark. This research is important as it addresses the problem of ‘hallucinations’ and ‘freshness’ in LLMs by enriching the prompt with additional factual and fresh data.
Publication date: 8 Feb 2024
Project Page: https://arxiv.org/abs/2402.05862
Paper: https://arxiv.org/pdf/2402.05862
