Knowledge Graphs are Implicit Reward Models: Path-Derived Signals Enable Compositional Reasoning

Researchers developed a new AI training method using knowledge graphs as reward models to improve compositional reasoning in specialized domains. The approach enables smaller 14B parameter models to outperform much larger frontier systems like GPT-5.2 and Gemini 3 Pro on complex multi-hop reasoning tasks in medicine.

• →Knowledge graphs can serve as implicit reward models to ground AI reasoning in verifiable domain facts.
• →The method uses supervised fine-tuning combined with reinforcement learning to train models on short reasoning paths that generalize to complex queries.
• →A 14B parameter model trained with this approach outperformed GPT-5.2 and Gemini 3 Pro on difficult medical reasoning tasks.
• →Path-derived rewards encourage models to compose intermediate axioms rather than just optimizing final answers.
• →The approach demonstrates robustness against adversarial perturbations and option-shuffling stress tests.