PathRouter: Aligning Rewards with Retrieval Quality in Agentic Graph Retrieval-Augmented Generation

PathRouter is a new training framework for agentic Graph Retrieval-Augmented Generation (GraphRAG) that addresses issues of reward aliasing and search-update ambiguity in reinforcement learning. By evaluating trajectories based on both answer correctness and evidence-path overlap, PathRouter reduces reliance on shortcuts while enhancing evidence-seeking behavior. Experimental results demonstrate that PathRouter improves answer F1 scores by an average of 3.1 for 3B models and 4.9 for 7B models across six QA benchmarks, making it a significant advancement for practitioners focused on optimizing retrieval quality in LLM applications.