Coupled Graph–Policy Distillation for Personalized Medication Safety in Older Adults with Multimorbidity

Older adults with multimorbidity often face complex medication regimens where clinical guidelines conflict or depend on specific patient factors (e.g., geriatric risks, unreported conditions). Standard large language models (LLMs) often fail to identify these missing, safety-critical variables, leading to potentially harmful recommendations. This paper asks: How can an AI agent effectively identify missing safety information, update its reasoning, and provide evidence-based medication plans in an interactive clinical setting?

The authors introduce ATLAS, a multi-agent framework that treats medication review as an iterative, evidence-based consultation. ATLAS utilizes a three-layer architecture:

1. Orchestration and Context Layer: Grounds the patient state and manages the dialogue.
2. Graph Personalization and Safety Audit Layer: Distills relevant clinical guidelines into a Patient-Specific Medication Conflict Graph (PMCG), which is updated as new information is acquired.
3. Decision Synthesis and Verification Layer: Executes a symbolic, risk-first policy to screen for contraindications, assess cautions, and verify the final plan.

ATLAS does not rely on a fixed questionnaire; instead, it uses the PMCG to identify unresolved safety dependencies and generates targeted questions to resolve them. The authors also introduce GeriMedBench, an interactive benchmark designed to evaluate an agent's ability to acquire missing information under a limited query budget and revise its decisions accordingly.

ATLAS significantly outperforms existing proprietary and open-source LLM baselines across multiple benchmarks. On the European non-interactive multimorbidity benchmark, it achieved a 92.04% Strict Success Rate, exceeding the strongest baseline by 53.73 points. In the interactive Asian multimorbidity benchmark (GeriMedBench), ATLAS demonstrated superior decision revision and safety, with no unsafe recommendations. Blinded clinician evaluations confirmed that ATLAS provided higher-quality, safer, and more complete medication advice compared to leading models like Gemini 3.1 Pro.