AgentKGV framework improves knowledge graph fact verification with two-stage training and dynamic routing

Knowledge graphs (KGs) constructed automatically from large-scale corpora often contain factual errors due to noisy sources and extraction failures, making reliable verification at industrial scale a persistent challenge. A new arXiv preprint proposes AgentKGV, an agentic LLM-RAG framework designed to address this by integrating dynamic routing and iterative query rewriting to handle surface-form mismatches in document-level retrieval.

The framework introduces a two-stage training strategy to improve both accuracy and cost-efficiency. First, turn-level distillation-based supervised fine-tuning (SFT) transfers reasoning ability from a larger teacher model to a smaller model, stabilizing query rewriting and reasoning. Second, trajectory-level GRPO optimizes the search policy to reduce unnecessary retrieval calls without sacrificing accuracy.

On the long-tail-predicate split of the open-domain T-REx benchmark, AgentKGV improves macro-F1 over single-turn RAG by 5.5 percentage points. When combined with the two-stage training strategy, performance improves by an additional 9.4 percentage points. The framework also reduces the average number of search calls from 3.24 to 1.63 without lowering accuracy.

The authors—Yumin Heo, Hyeon-gu Lee, Sumin Seo, and Youngjoong Ko—position AgentKGV as a step toward more reliable and scalable verification of automatically constructed knowledge graphs, which are foundational to many industrial AI systems.