AutoGraph-R1
About this tool
Official code for an ACL 2026 paper treating model-based knowledge-graph construction as reinforcement learning. Rewards reflect a graph's usefulness in downstream RAG Q&A, using graphs as either knowledge carriers or indexes. VeRL- and vLLM-based training and evaluation scripts support reproducing and extending task-driven GraphRAG research.
- Feature 1: End-to-end reinforcement learning rewards downstream RAG answer quality.
- Feature 2: Task-aware rewards for graphs as knowledge carriers or indexes.
- Feature 3: VeRL and vLLM training, with FSDP-to-Hugging Face checkpoint conversion.
- Feature 4: Inference and benchmark reproduction scripts, plus evaluation of custom graphs.
Resource information
- Maintainer
- HKUST-KnowComp, Associate Professor Yangqiu Song's group.
- Type
- Tool / library
- Category
- Knowledge graphs & graph computing
- Runs on
- Tool or library
- Who it's for
- Research teams studying GraphRAG, graph construction and model reinforcement learning, especially comparisons on multihop Q&A benchmarks.
- License at time of listing
- MIT
- Source
- GitHub
