World models of agent societies

Core objective

A central long-term goal of imec.AI-labs is to develop the ability to model complex societies of interacting agents that communicate in natural language. The aim is to predict how these agent societies will evolve over time by capturing how distributed groups of agents exchange information, update beliefs, react to external conditions, and collectively shape future outcomes.

Broad applications

By successfully learning “World Models” of agent societies, researchers aim to unlock capabilities across several domains:

• Forecasting dynamics: predicting large-scale economic and social trajectories.
• System design: designing multi-agent AI systems capable of coordinating, negotiating, and making collective decisions.
• Simulation and counterfactual reasoning: creating robust simulation environments to study policy interventions, governance, and technological shifts. This includes exploring how alternative communication patterns, incentive structures, or environmental pressures might alter a society’s long-term path.

Ultimately, this research aims to bridge the gap between micro-level agent behaviors and macro-level emergent phenomena, providing a principled framework to understand and steer complex sociotechnical systems.

Primary Use-Case: Modeling Hospitals and Patients

To ground this theoretical research in a real-world application, the project currently focuses on the healthcare domain: forecasting the longitudinal future of patients through the prediction of medical reports. In this scenario, clinicians, patients, and medical systems represent the interacting agents, producing a structured stream of natural-language reports over time.

Overcoming Traditional AI Limitations

Traditional healthcare AI systems typically rely on highly specialized models tailored to narrow outcomes, such as predicting readmission risk, or specialized modalities like imaging and genomics. While effective within their domains, these models are often brittle, require costly datasets, and struggle to generalize across different populations, institutions, or medical specialties.

The Generative Forecasting Approach

Instead of narrow predictive pipelines, the project aims to build a general-purpose foundation model. * Medical Reports as an Abstraction Layer: Medical reports, such as progress notes, radiology reports, and discharge summaries, encode the observations, reasoning, and plans of clinicians. Because they are ubiquitous across institutions and diseases, they provide a scalable, domain-agnostic interface for longitudinal prediction. * Temporal Sequence Modeling: A patient is modeled as a temporal sequence of medical reports written by different doctors. Forecasting is treated as a generative modeling problem: predicting the next medical report (or distribution of future reports) conditioned on historical reports and optional constraints like medical specialty or planned interventions.

Unlocked Capabilities in Healthcare

By conceptualizing patient evolution as the generation of future medical records, the system enables several advanced capabilities: 1. Longitudinal trajectory prediction: Generating future clinical narratives that describe the expected evolution of conditions and treatments. 2. Specialty-oriented forecasting: Constraining the AI’s predictions to specific specialties to analyze focused aspects of patient health. 3. Treatment simulation: Conditioning the generated reports on specific medications or interventions to explore counterfactual, in silico scenarios. 4. Care-pathway analysis: Examining how predictions differ under various clinician profiles or referral patterns. 5. Outcome-oriented optimization: Utilizing reinforcement learning or preference modeling to identify interventions that steer patient trajectories toward desirable futures.

By utilizing medical reports as a dynamic temporal signal, this healthcare use-case serves as an ideal proving ground for the broader ambition of forecasting complex, interacting agent societies.