Agentifying Patient Dynamics within LLMs through Interacting with Clinical World Model
TLDR
SepsisAgent uses a learned Clinical World Model and a propose-simulate-refine workflow to improve LLM-based sepsis treatment decisions, outperforming baselines on MIMIC-IV.
Reasoning
The paper presents a novel integration of a world model with LLM agents for clinical decision-making, with strong empirical results on a real-world dataset. However, the abstract lacks details on the world model architecture and generalizability beyond sepsis.
Read-first score
Read-first score 61.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.
Field roles
Rank sensitivity
Stability: volatile; rank range: 398.
Keyword Scores
Deep Analysis
Innovations
- World model-augmented LLM agent for sepsis treatment recommendation
- Propose-simulate-refine workflow before committing to a prescription
- Three-stage curriculum training: patient-dynamics SFT, propose-simulate-refine behavior cloning, and world-model-based agentic reinforcement learning
- Demonstration that repeated interaction with the Clinical World Model enables learning of patient evolution regularities useful even without simulator access
Methodology
SepsisAgent uses a learned Clinical World Model to simulate patient responses under candidate fluid–vasopressor interventions and follows a propose–simulate–refine workflow. Training is conducted via a three-stage curriculum: patient-dynamics supervised fine-tuning, propose–simulate–refine behavior cloning, and world-model-based agentic reinforcement learning. Evaluation is performed on MIMIC-IV sepsis trajectories against traditional RL and LLM-based baselines using off-policy value, guideline adherence, and unsafe-action metrics.
Key Results
SepsisAgent outperforms all traditional RL and LLM-based baselines in off-policy value while achieving the best safety profile under guideline adherence and unsafe-action metrics.