KG-R1: Reinforcement Learning Optimizes Knowledge Graph RAG for LLMs
Researchers introduce KG-R1, an agentic framework that uses reinforcement learning to optimize knowledge-graph retrieval-augmented generation (KG-RAG) for large language models. Unlike traditional KG-RAG systems that rely on fixed pipelines of multiple LLM modules (planning, reasoning, responding), KG-R1 employs a single agent that interacts with knowledge graphs as its environment, learning to retrieve information step by step and integrating it into reasoning and generation. This unified process reduces inference costs and improves transferability across different graph schemas. Evaluated on Knowledge-Graph Question Answering (KGQA) benchmarks, KG-R1 using Qwen 2.5-3B achieves higher answer accuracy with fewer generation tokens compared to existing methods. The work addresses key limitations of current KG-RAG systems, including high inference costs and schema-specific performance. The paper is available on arXiv under reference 2509.26383.
Key facts
- KG-R1 is an agentic framework for KG-RAG using reinforcement learning.
- It uses a single agent instead of multiple LLM modules.
- The agent interacts with KGs as its environment.
- It learns to retrieve information step by step.
- It integrates retrieval, reasoning, and generation in a unified process.
- Evaluated on KGQA benchmarks.
- Uses Qwen 2.5-3B model.
- Improves answer accuracy with fewer generation tokens.
- Addresses high inference costs and schema-specific performance.
- Paper available on arXiv: 2509.26383.
Entities
Institutions
- arXiv