graph-native-reasoning

This paper introduces Graph-PRefLexOR, a family of graph-native reasoning models fine-tuned with group relative policy optimization (GRPO) to generate traceable scientific hypotheses through explicit reasoning phases. The method achieves 40-65% improvements over base models in reasoning traceability and demonstrates enhanced semantic diversity and conceptual recombination.