Proof-Augmented Retrieval and Reasoning: Supervising Language models for Knowledge Graph Completion with Link Predictors

ICLR 2026 Conference Submission3548 Authors

Keywords: Knowledge Graph Completion, Neurosymbolic

Abstract: We propose Proof-Augmented Retrieval and Reasoning (PARR), a Rewrite-Retrieve-Read framework that leverages interpretable link predictors to supervise the retrieval and reasoning of LLMs for Knowledge Graph Completion (KGC). We utilize proofs extracted from pre-trained interpretable link predictors to improve both retrieval and reasoning. First, we augment retrieval with extracted proofs for better sub-graph retrieval. Second, we leverage proofs as ground-truth retrievals to train a Rewriter LLM for query rewriting. Finally, we utilize the proofs as chain-of-thought data to train a Reasoner LLM for link prediction given the retrieved context. We evaluate our framework on FB15K-237 and WN18RR under both transductive and inductive settings, demonstrating competitive performance compared to state-of-the-arts.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 3548