RLKGC: Reinforcement Learning Retrieval with Large Language Models for Knowledge Graph Completion

Inductive knowledge graph completion (KGC) aims to predict missing facts that involve unseen entities. Many recent methods rely strongly on pre-computed reasoning paths or dense graph structure, making them unreliable when paths are sparse or noisy. At the same time, type information and local neighborhood facts often provide useful clues, but they are not fully exploited by existing models. We present RLKGC, a framework that combines policy-guided multi-hop exploration with lightweight type and subgraph reasoning using a language model. The reinforcement learning (RL) module searches for meaningful paths and is trained with reward shaping and action dropout to reduce false-negative signals and avoid spurious shortcuts. The type module checks whether a candidate aligns with the implicit type pattern required by the query relation, and the subgraph module uses RL paths together with nearby facts to assess triple plausibility. Across three standard KGC benchmarks, RLKGC improves inductive and few-shot prediction quality, achieving gains of 4.78% in MRR and 7.82% in Hits@1, and outperforming prior methods in 9 of 12 evaluation settings.