RANGER: Retrieval Agent for Navigating Graphs with Semantic Evaluation Rewards
- Published
- Source
- arXiv
- Paper number
- 1367
- Field
- RAG / Agents
- arXiv ID
- 2610.28010
Key points
- The LLM served as a mentor and evaluated the semantic relevance of paths found by the RL agent as Success, Highly Relevant, or Failure, giving rewards of 1, 0.5, and 0 respectively.
- A Transformer-based policy agent trained with GRPO directly navigates the knowledge graph and gathers triplets needed for the query.
- In standalone retrieval on HotpotQA, RANGER with the LLM reward improved the LLM-as-a-Judge score by approximately 3.6% over WaterCircles and by approximately 3.0% over GRPO with the heuristic reward.
- In hybrid retrieval, it improved the LLM Judge score over WaterCircles by approximately 20.4% on HotpotQA, 4.4% on TriviaQA, and 7.7% on DiaASQ.
- Reducing the mentor LLM from Mistral Small 3.2 (24B) to Ministral 3 (8B) dropped performance below the WaterCircles baseline.
- The agent was trained in a self-supervised manner by automatically generating question-answer pairs from the knowledge graph itself, without human-labeled data.
Paper links
External research summaries. These are not HDATF publications or measured product results.