Kinship Reasoning as a Testbed for Neuro-Symbolic Architectures: Comparing LLM–Prolog Integration Strategies and Assessing Cultural Bias

Large language models (LLMs) have demonstrated impressive capabilities across a wide range of tasks, yet relational and multi-hop reasoning remain challenging. Neuro-symbolic approaches seek to address these limitations by combining the linguistic competence of LLMs with the precision and transparency of symbolic reasoning systems. In this paper, we use kinship reasoning as a controlled testbed for evaluating alternative LLM–Prolog integration strategies. Using the KinshipQA benchmark, a synthetic dataset generator capable of producing culturally grounded family trees and kinship questions with known symbolic ground truth, we compare several neuro-symbolic architectures that differ in the division of labour between the language model and the Prolog inference engine. Our first experiment compares alternative architectures ranging from a pure LLM baseline to systems that employ LLM-generated Prolog kinship theories. The results identify the most effective integration strategy and suggest that LLMs are more reliable as generators of reusable symbolic theories than as online logical reasoners. In a second experiment, we use the architecture requiring the least domain-specific input to compare performance across alternative kinship classification systems, including Eskimo and Crow kinship. By evaluating not only question-answering accuracy but also the correctness of the generated Prolog theories, we show how neuro-symbolic representations make differences in culturally specific knowledge explicit and measurable. The proposed framework therefore provides both a methodology for evaluating neuro-symbolic architectures and a means of analysing the impact of cultural variation in kinship classification on reasoning performance.