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

2026-07-10

AI Summary

This work addresses the challenge of efficiently and accurately verifying facts in automatically constructed knowledge graphs, which are often corrupted by noise and extraction errors—particularly at industrial scale. The authors propose AgentKGV, a novel framework that integrates an LLM-RAG architecture with dynamic routing and iterative query rewriting to mitigate surface-form mismatches in document retrieval. A key innovation is its two-stage training strategy: first, distilled supervised fine-tuning (SFT) enables stable reasoning with a smaller model; second, trajectory-level generalized reinforcement policy optimization (GRPO) refines the search policy to reduce redundant retrievals. Experiments on the long-tail predicate split of T-REx show that AgentKGV improves macro F1 by 14.9 points over single-turn RAG, while GRPO cuts the average number of retrievals from 3.24 to 1.63 without sacrificing accuracy.

Abstract

Knowledge graphs (KGs) are often automatically constructed from large-scale corpora, but they inevitably contain factual errors due to noisy sources and extraction failures, and verifying them reliably at industrial scale remains a critical challenge. To address this, we propose AgentKGV, the Agentic LLM-RAG framework for KG fact Verification, that integrates dynamic routing and iterative query rewriting, which handles surface-form mismatch in document-level retrieval. To make this framework more accurate and cost-efficient for industrial deployment, we further introduce a two-stage training strategy: turn-level distillation-based SFT that transfers reasoning ability from a large teacher model into a small model for stable query rewriting and reasoning, and trajectory-level GRPO that optimizes the search policy to reduce unnecessary retrieval at scale. On the long-tail-predicate split of the open-domain T-REx benchmark, our framework improves macro-F1 over single-turn RAG by 5.5 \%p, and two-stage training does it further by 9.4 \%p. GRPO also cuts the average number of search calls from 3.24 to 1.63 without lowering accuracy.

Problem

• Knowledge Graph
• Fact Verification
• Industrial Scale
• Factual Errors
• Large-scale Corpora

Innovation

• Agentic RAG
• Two-stage training
• Query rewriting
• Fact verification
• Knowledge graph