REPANA: Reasoning Path Navigated Program Induction for Universally Reasoning over Heterogeneous Knowledge Bases

Jinxin Liu, Shulin Cao, Jiajie Zhang, Ao Zou, Xiaoyin Che, Liangming Pan, Lei Hou, Juanzi Li

Keywords: Knowledge Base QA; Low Resource Reasoning; Multi-hop QA; Reasoning Interpretability;

Abstract: Program induction is a typical approach that helps Large Language Models (LLMs) in complex knowledge-intensive question answering over knowledge bases (KBs) to alleviate the hallucination of LLMs. However, the accurate program induction usually requires a large number of high-quality parallel data of a specific KB, which is difficult to acquire for many low-resource KBs. Additionally, due to heterogeneity of questions and KB schemas, the transferability of a model trained on a single dataset is poor. To this end, we propose REPANA, a reasoning path navigated program induction framework that enables LLMs to reason over heterogeneous KBs. We decouple the program generation capability into perceiving the KB and mapping questions to program sketches. Accordingly, our framework consists of two main components. The first is an LLM-based navigator, which retrieves reasoning paths of the input question from the given KB. The second is a KB-agnostic parser trained on data from multiple heterogeneous datasets, taking the navigator's retrieved paths and the question as input and generating the corresponding program. Experiments show that REPANA exhibits strong generalization and transferability. It can directly perform inference on datasets not seen during training, outperforming other SoTA low-resource methods and even approaching the performance of supervised methods.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 13193