PathISE: Learning Informative Path Supervision for Knowledge Graph Question Answering

Researchers have introduced PathISE, a novel framework designed to enhance Knowledge Graph Question Answering (KGQA) by learning high-quality intermediate supervision from answer-level labels. Current KGQA methods often rely on retrieval-augmented generation but struggle with the high cost and resource intensity of obtaining question-relevant paths or subgraphs for training. PathISE addresses this by employing a lightweight transformer-based estimator to assess the informativeness of relation paths, thereby constructing pseudo path-level supervision. This supervision is distilled into a Large Language Model (LLM) path generator, which produces grounded paths for compact evidence in inductive answer reasoning. Extensive experiments across three KGQA benchmarks demonstrate that PathISE achieves competitive or state-of-the-art performance. Crucially, it provides reusable supervision signals that can improve existing models without depending on expensive LLM-refined data. The study highlights a significant advancement in making KGQA systems more efficient and scalable by reducing reliance on costly manual or computational supervision resources.