<ul><li>Teaching Large Language Models (LLMs) to reason on graphs using Reinforcement Learning (RL).</li><li>Introducing G1 approach demonstrating significant improvement in LLMs' graph reasoning abilities through RL training on synthetic graph-theoretic tasks.</li><li>Curated Erdős dataset comprising diverse graph tasks used for RL training, resulting in substantial enhancements in graph reasoning.</li><li>RL-trained models exhibit strong generalization to unseen tasks, domains, and graph encoding schemes, indicating potential for building strong graph reasoners by finetuning LLMs with RL.</li></ul>

G1: Teaching LLMs to Reason on Graphs with Reinforcement Learning

Discover more