GraphRAG-IRL: Personalized Recommendation with Graph-Grounded Inverse Reinforcement Learning and LLM Re-ranking

A team of researchers has published a new paper detailing GraphRAG-IRL, a hybrid AI framework designed to overcome the significant limitations of using large language models (LLMs) as standalone recommendation engines. The core problem is that while LLMs are excellent semantic reasoners, they are unreliable rankers, suffering from poor calibration, sensitivity to the order of candidate items, and a tendency to amplify popularity bias. GraphRAG-IRL tackles this by creating a robust, multi-stage pipeline that first builds a heterogeneous knowledge graph (GraphRAG) connecting items, categories, and concepts to capture rich contextual signals.

This graph-grounded data feeds into a Maximum Entropy Inverse Reinforcement Learning (IRL) model, which learns user preferences from their sequential behavior to produce a calibrated, pre-ranked candidate list. The key innovation is that the computationally expensive LLM is only applied to this short, high-quality list. The LLM acts as a final re-ranker, using persona-guided prompts to make nuanced semantic judgments, which are then fused with the IRL scores. This division of labor leverages the strengths of each component: the IRL model provides robust, behavior-based ranking, while the LLM adds high-level semantic understanding.

Experiments on major datasets like MovieLens and KuaiRand demonstrate the framework's effectiveness. The GraphRAG-enhanced IRL model alone improved recommendation accuracy (NDCG@10) by 15.7% and 16.6% over standard supervised baselines. The researchers found the combination of GraphRAG and IRL to be superadditive, meaning their joint improvement exceeded the sum of their individual gains. Adding the final LLM fusion stage pushed performance even further, yielding up to a 16.8% NDCG@10 improvement over the IRL-only baseline on MovieLens, with consistent 4-6% gains across different LLM providers on KuaiRand.