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

2026-07-10 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

AgentKGV is an Agentic LLM-RAG framework designed for the fact verification of knowledge graphs (KGs), addressing factual errors arising from noisy sources and extraction failures at industrial scale. The framework integrates dynamic routing and iterative query rewriting to effectively handle surface-form mismatches during document-level retrieval. To enhance accuracy and cost-efficiency for industrial deployment, AgentKGV employs a two-stage training strategy. This includes turn-level distillation-based Supervised Fine-Tuning (SFT) to transfer reasoning abilities from a large teacher model to a smaller one for stable query rewriting, and trajectory-level Generalized Policy Optimization (GRPO) to optimize the search policy and reduce unnecessary retrieval. On the long-tail-predicate split of the T-REx benchmark, AgentKGV improves macro-F1 over single-turn RAG by 5.5 %p, with the two-stage training adding a further 9.4 %p improvement. GRPO specifically reduces average search calls from 3.24 to 1.63 without accuracy loss.

Key takeaway

For Machine Learning Engineers building knowledge graph verification systems, AgentKGV offers a robust approach to improve factual accuracy and operational efficiency. You should consider integrating agentic LLM-RAG with dynamic routing and iterative query rewriting to handle complex retrieval challenges. Furthermore, adopting a two-stage training strategy, particularly GRPO, can significantly reduce inference costs by cutting search calls by nearly half without compromising verification accuracy.

Key insights

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

Principles

• Iterative query rewriting improves retrieval.
• Distillation transfers reasoning to smaller models.
• Policy optimization reduces search overhead.

Method

AgentKGV combines dynamic routing and iterative query rewriting for KG fact verification. Its two-stage training uses distillation-based SFT and trajectory-level GRPO to optimize reasoning and retrieval.

In practice

• Implement agentic RAG for KG verification.
• Apply two-stage training for efficiency.
• Use GRPO to cut retrieval calls.

Topics

• Knowledge Graphs
• Fact Verification
• LLM-RAG Frameworks
• Agentic AI
• Two-Stage Training
• Policy Optimization

Best for: Research Scientist, AI Engineer, 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 efficiency.
• 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.
• Knowledge Layers: A Practical Fix for Stale RAG in Agentic AI — Pre-building policy-enforced knowledge layers before agent execution resolves stale RAG issues in multi-step AI agents.
• RAG Has Evolved into Multi-modal AI Search Agents (RL, HaPO) — Multimodal AI search agents utilize knowledge graphs and checkpoint-based RL for efficient multi-step visual reasoning.
• Bringing Agentic Search to Earth Observation Data Discovery — Agentic search, amplified by knowledge graphs and LLM reranking, significantly improves Earth Observation data discovery.
• 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.

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 Computation and Language.