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

2026-07-13 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, extended

Summary

AgentKGV is an Agentic LLM-RAG framework designed for fact verification in knowledge graphs (KGs), addressing factual errors from automatic construction. It integrates dynamic routing and iterative query rewriting to handle surface-form mismatches in document retrieval. To enhance accuracy and cost-efficiency for industrial deployment, AgentKGV employs a two-stage training strategy: turn-level distillation-based SFT for stable query rewriting and reasoning, and trajectory-level GRPO to optimize search policy and reduce unnecessary retrieval. On the T-REx long-tail-predicate split, AgentKGV improves macro-F1 by 5.5 %p over single-turn RAG, with two-stage training adding another 9.4 %p. GRPO also reduces average search calls from 3.24 to 1.63 without accuracy loss, demonstrating its effectiveness on both English open-domain and Korean enterprise KGs.

Key takeaway

For Machine Learning Engineers building robust KG fact verification systems, AgentKGV's two-stage training offers a path to significantly improve accuracy and reduce inference costs. You should consider implementing distillation-based SFT to stabilize query rewriting for domain-specific predicates and GRPO to optimize search policies, cutting average retrieval calls by nearly half while maintaining or improving F1 scores. This approach is particularly valuable for industrial-scale KGs with long-tail predicates.

Key insights

AgentKGV uses agentic LLM-RAG with two-stage training to verify knowledge graph facts, improving accuracy and efficiency.

Principles

• Dynamic routing avoids unnecessary retrieval calls.
• Iterative query rewriting handles modality gaps.
• Distillation transfers reasoning to smaller models.

Method

AgentKGV dynamically routes verification, then iteratively rewrites compressed KG triples into natural language queries for document retrieval. A two-stage training strategy uses distillation-based SFT for query rewriting and GRPO for search policy optimization.

In practice

• Apply dynamic routing to reduce LLM inference costs.
• Use iterative query rewriting for complex retrieval.
• Distill large model knowledge to smaller models.

Topics

• Knowledge Graph Verification
• Agentic LLMs
• Retrieval-Augmented Generation
• Two-Stage Training
• Supervised Fine-Tuning
• Reinforcement Learning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• AgentKGV: Agentic LLM-RAG Framework with Two-Stage Training for the Fact Verification of Knowledge Graphs — AgentKGV uses agentic LLM-RAG with two-stage training to verify knowledge graph facts, improving accuracy and reducing retrieval costs.
• Multi-Sourced, Multi-Agent Evidence Retrieval for Fact-Checking — WKGFC unifies structured KGs and web evidence via an LLM agent for adaptive, self-improving fact-checking.
• Scaling AI Knowledge Systems: Lessons from the DHTMLX MCP Server — Scaling RAG systems for multiple distinct products requires intelligent query routing and optimized knowledge structuring.
• Code-on-Graph: Iterative Programmatic Reasoning via Large Language Models on Knowledge Graphs — Code-on-Graph integrates LLMs with KGs using programmatic reasoning, representing schemas as Python classes for flexible, scalable knowledge utilization.
• TechGraphRAG: An Agentic Graph-Augmented RAG Framework for Technical Literature Reasoning — Agentic RAG frameworks can enhance technical reasoning by integrating multi-step pipelines, knowledge graphs, and external search for robust evidence grounding.
• Data Analysis in the Wild: Benchmarking Large Language Models Against Real-World Data Complexities — Current LLMs, even with agentic support, significantly underperform on real-world complex data analysis tasks.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.