The authors propose Neurosymbolic Alignment, a training-time framework that couples a 7B clinical LLM with an HGNN-based Physiological World Model built over an 847K-node biomedical . Candidate responses are scored using homeostatic constraints, multi-hop path plausibility, and drug-interaction penalties, then these scores drive iterative on-policy ORPO updates.
On the Clinical Safety Benchmark (CSB), the method improves Clinical Safety Score (CSS) from 69.5% to 90.8%, reduces physician-evaluated Harm Rate from 14.1% to 5.1%, and improves Drug Interaction Detection from 72.8% to 91.6%. These gains outperform GPT-4 (5-shot) and an inference-time self-correction pipeline, and are corroborated by an independent rule-engine safety score. The approach is interesting because it demonstrates that grounding in structured physiological knowledge produces measurable safety improvements without needing the graph at inference time.