AgentKGV boosts KG fact verification with 14.9% F1 gain and 50% fewer searches

AgentKGV is a new framework from researchers at Dong-A University that addresses a critical bottleneck in industrial-scale knowledge graph (KG) maintenance: verifying factual accuracy in automatically constructed KGs, which are prone to errors from noisy sources and extraction failures. The framework combines an agentic LLM with a retrieval-augmented generation (RAG) pipeline that dynamically routes queries and iteratively rewrites them to overcome surface-form mismatches between KG facts and free-text evidence. This design ensures document-level retrieval is both precise and cost-effective.

To make the system practical for deployment, the authors introduce a two-stage training strategy. First, turn-level supervised fine-tuning (SFT) with distillation transfers reasoning capabilities from a large teacher model to a smaller, faster student model for stable query rewriting and multi-step reasoning. Second, trajectory-level GRPO (a policy optimization technique) learns to minimize unnecessary retrieval steps, effectively reducing the average number of search API calls from 3.24 to 1.63—a ~50% reduction—while maintaining or improving accuracy. On the long-tail predicate split of the open-domain T-REx benchmark, AgentKGV boosts macro-F1 by 5.5 percentage points over a single-turn RAG baseline, and the two-stage training adds another 9.4 points, for a total gain of 14.9%.