AgentKGV: Agentic LLM-RAG Framework with Two-Stage Training for the Fact Verification of Knowledge Graphs

Short summary

AgentKGV is an agentic LLM-RAG framework for verifying facts in automatically constructed knowledge graphs, using dynamic routing and iterative query rewriting to handle retrieval mismatches. A two-stage training strategy combines turn-level distillation-based SFT for stable reasoning and trajectory-level GRPO to optimize search policy, cutting retrieval calls from 3.24 to 1.63. On the T-REx benchmark, the framework improves macro-F1 by 5.5 points over single-turn RAG, with two-stage training adding another 9.4 points.

• •Agentic LLM-RAG framework for KG fact verification with dynamic routing and query rewriting
• •Two-stage training: distillation-based SFT for reasoning + GRPO for search policy optimization
• •Improves macro-F1 by 15 points total over single-turn RAG while halving search calls

Generated with AI, which can make mistakes.