LLMs vs. World Models — And Why World Models Fit Healthcare Better
— by Vivax
Large language models predict the next word. World models predict what happens next in the real world.
A large language model (LLM) learns the statistical patterns of human text. Trained to predict the next word across billions of documents, it becomes remarkably fluent — but its knowledge is a map of how words co-occur, not of how the world actually works. A world model is a different kind of system: it learns an internal, predictive simulation of an environment, so it can anticipate what happens next and plan before acting. The idea was formalized in Ha and Schmidhuber's 2018 paper 'World Models,' and championed by Yann LeCun as the path toward machines that genuinely understand cause and effect.
That distinction matters most where mistakes are costly. Because an LLM has no grounded model of the body, a disease, or a treatment, it can produce confident, fluent, and completely wrong answers — what researchers call medical hallucinations. A 2025 study cataloguing hallucinations in foundation models warns that fabricated dosages, drug interactions, or diagnostic criteria can lead directly to harm, and that these errors are hard to catch precisely because they sound authoritative. Fluency is not understanding, and in the clinic that gap can be dangerous.
World models are gaining ground in medicine for exactly the reasons LLMs struggle. Clinicians reason counterfactually all day — what happens if we wait, escalate, or change the drug? A world model learns multimodal, temporally coherent, action-conditioned representations: it can simulate how a patient's state evolves under different treatment plans and 'imagine' outcomes before they happen. A 2025 survey, 'Beyond Generative AI: World Models for Clinical Prediction, Counterfactuals, and Planning,' argues that this causal, predictive structure is what grounded clinical decision support actually needs — something scaling language models alone has not delivered.
Medicine is also physical. Surgery, imaging, robotics, and connected devices all unfold in space and time, governed by the laws of the real world. NVIDIA frames this as the shift from generative AI to physical AI, with world foundation models like Cosmos learning to generate physically accurate scenarios — a digital twin of reality that robots and surgical platforms can reason over safely. A clinical AI that understands both the patient and the room around them needs this shared, grounded picture, not disconnected text predictions.
This is the bet behind Vivax. We are building toward a unified clinical world model — one that understands the patient (physiology, history, context) and the physical world around them (the operating room, instruments, imaging, and devices) — as the foundation of Acudx and every device we ship. LLMs remain powerful tools for language; but for safe, grounded, decision-grade medical AI, world models are the more suitable foundation. The references below are a good place to start reading.