AgentKGV uses two-stage training to improve KG fact verification and reduce search calls

Researchers propose AgentKGV, an agentic LLM-RAG framework designed for the reliable fact verification of knowledge graphs at industrial scale. The system integrates dynamic routing and iterative query rewriting to address surface-form mismatch in document-level retrieval.

• The framework employs a two-stage training strategy: turn-level distillation-based SFT for stable reasoning and trajectory-level GRPO to optimize search policy.
• On the long-tail-predicate split of the T-REx benchmark, AgentKGV improves macro-F1 over single-turn RAG by 5.5%.
• The two-stage training provides an additional 9.4% improvement in macro-F1.
• GRPO reduces the average number of search calls from 3.24 to 1.63 without lowering accuracy.

This approach enables more accurate and cost-efficient verification by reducing unnecessary retrieval while maintaining high performance on noisy knowledge graphs.