PolicyRAG: Prompt-Guided Symbolic Graph Memory for Interpretable Multi-Hop Retrieval

ICLR 2026 Conference Submission18492 Authors

Keywords: Large Language Models, Retrieval Augmented Generation, Graph Based Reasoning, Personalized PageRank, Knowledge Graphs, Symbolic Graph Memory, Multi-Hop Question Answering

Abstract: Retrieval augmented generation is a powerful way to ground large language models in external knowledge, yet most pipelines still treat the prompt as instructions for text production rather than as a control surface for retrieval. We introduce PolicyRAG, a framework that recasts retrieval as an explicit, auditable policy operating over a symbolic graph memory. Text is organized into lightweight typed links between entities and passages, enabling transparent search and controllable evidence selection. Beyond accuracy, the policy is compact and human editable, supporting governance, domain adaptation, and safety review without retraining. At query time, the policy seeds candidate entities, invokes brief LLM calls only for disambiguation and local gating, and performs symbolic traversal with Personalized PageRank (PPR). The resulting scores are projected to passages and finalized with a small, transparent re-ranker, producing a per-query trace of seeds, paths, and scores for explainable evidence selection. Compared with long-context expansion, the policy keeps test-time compute modest while preserving answer quality. On multi-hop question answering benchmarks, PolicyRAG achieves state-of-the-art results on HotpotQA (F1 80.7), 2WikiMultiHopQA (F1 78.9), and MuSiQue (F1 55.9) while remaining fully auditable and training free. We also assess domain adaptability on domain-specific datasets. By coupling symbolic structure with prompt level control, PolicyRAG provides a practical route from question to verifiable evidence and advances accuracy, efficiency, and trust in retrieval augmented generation.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 18492