Source: arXiv — 2026-07-10

Summary

AgentKGV is a two-stage-trained agentic RAG framework for verifying facts against knowledge graphs, combining turn-level distillation supervised fine-tuning with trajectory-level GRPO reinforcement learning, plus dynamic query rewriting and routing. On a long-tail split of the T-REx benchmark, it improves macro-F1 by 5.5 points over single-turn RAG, gains another 9.4 points from the two-stage training regimen, and GRPO training cuts the average number of search calls per query from 3.24 to 1.63 without losing accuracy.

Key Takeaways

• Two-stage training: turn-level distillation SFT teaches the agent good single-step behavior, then trajectory-level GRPO optimizes the whole multi-step search trajectory.
• +5.5 macro-F1 points over single-turn RAG baselines on a long-tail T-REx fact-verification split.
• An additional +9.4 points attributable specifically to the two-stage training approach versus single-stage alternatives.
• GRPO training nearly halves search cost: average search calls per query drop from 3.24 to 1.63, with no accuracy loss — a direct compute/latency win.
• Dynamic query rewriting and routing let the agent adapt its search strategy per query rather than following a fixed retrieval pattern.

Reel Script

Hook Most RAG agents either search too little and hallucinate, or search over and over to be safe, burning latency and cost. A new paper trains an agent that gets more accurate while cutting its own search calls almost in half.

Core Concept RAG, retrieval-augmented generation, means instead of trusting the model's memory, it looks facts up — here, specifically against a knowledge graph, a structured database of entities and their relationships. The problem: agentic RAG systems that decide for themselves when and what to search often over-search, treating every query the same. AgentKGV trains the agent in two stages: first, distillation SFT, a turn-level lesson where the agent learns from good example search steps, like a student copying a teacher's individual moves. Then trajectory-level GRPO, a reinforcement-learning stage that grades the whole multi-step search trajectory as one unit, rewarding efficient paths to a correct answer, not just correct individual steps. That second stage is what teaches the agent to stop over-searching.

Hands-On The numbers, on a long-tail split of the T-REx fact-verification benchmark: baseline single-turn RAG gets a macro-F1 score, agentic multi-turn search adds 5.5 points on top of that, and the two-stage training regimen adds another 9.4 points beyond a single-stage approach. But the number that matters for anyone running these in production: GRPO training drops the average search calls per query from 3.24 down to 1.63, search cost nearly cut in half, while accuracy holds steady. That's the kind of result that changes a RAG system's actual API bill and latency budget, not just its leaderboard score.

Takeaway My verdict: this is a genuinely useful pattern beyond just fact verification — training an agent to grade its whole search trajectory rather than each step in isolation is how you get RAG systems that are both more accurate and cheaper to run. If you're building agentic RAG, look at trajectory-level RL, not just SFT. Follow for more on what's actually moving RAG forward versus incremental noise.

Discussion

(No questions yet — ask follow-ups via a Claude Code chat session on this repo; answers get appended here.)