One Hop at a Time: Knowledge Graph-Grounded Curriculum Learning for Bottom-Up Clinical Reasoning

Abstract

Language models consistently struggle with compositional multi-hop reasoning in clinical medicine, in part because existing knowledge graphs expose only a limited set of relational patterns during training. We address this by scaling both the size and relational diversity of the underlying knowledge graph to construct a curriculum of multi-hop reasoning tasks. Beginning from the 9,958-node QA-GNN seed graph, we expand via UMLS data to construct a clinically comprehensive 45,746-node, 702,290-edge knowledge graph through an iterative pruning and bootstrapping methodology. We generate 116,859 question-answer pairs with chain-of-thought traces grounded in explicit relational paths derived from the KG. The resulting KG is evaluated through supervised fine-tuning of DeepSeek-R1-Distill-Qwen-14B and -7B models. The fine-tuned 14B model improves by +31.9 pp on MedQA and +9.71 pp on ICD-Bench over the base model, matching or exceeding Med3 on key benchmarks despite using fewer than half the parameters. Gains are larger on compositional three-hop chains containing a familiar training edge than on fully novel chains, indicating that the model learns reusable relational primitives rather than memorizing paths. This compositional benefit is capacity-dependent: the 7B model shows smaller gains, suggesting that exploiting relational structure requires sufficient scale. These findings establish that expanding the space of relational primitives is critical for enabling compositional reasoning in domain-specific language models. The complete KG and curriculum datasets are available at: \url{https://huggingface.co/datasets/jha-lab/med-si-dataset}, with trained model checkpoints linked in the description.