HyGRL Improves Multi-Entity Question Answering with Hybrid Graph Reasoning

Answering complex questions that involve multiple entities and require compositional reasoning presents a significant challenge for current retrieval-augmented language models (RAGs). Traditional RAG systems often lack dynamic reasoning capabilities, while standard Graph-RAGs are limited by structural sparsity. Furthermore, LLM-constructed Graph-RAGs can be prohibitively expensive. This paper introduces HyGRL, a novel, unified framework designed to overcome these limitations. HyGRL transforms unstructured text into structured knowledge graphs, creating a heterogeneous network that allows for more flexible and efficient evidence retrieval. This approach effectively merges the richness of textual information with the structured nature of knowledge graphs. The reasoning process within HyGRL is formulated as adaptive structure induction, learned through a robust two-stage process. Initially, imitation learning distills heuristic expert signals, which are then refined by reinforcement learning using LLM-driven preference rewards. Experiments demonstrate that HyGRL significantly outperforms state-of-the-art baselines in both answer accuracy and reasoning fidelity, all while maintaining low token costs and near real-time inference.