Research Update: Breakthrough in Population-Scale Simulations with Adaptive Agents

We're excited to announce that our paper "On the limits of agency in agent-based models" has been accepted as an Oral presentation at AAMAS 2025. Our research represents a fundamental breakthrough in agent-based modeling: enabling large language model (LLM) guided agents to scale from small simulations of hundreds to massive population-level simulations of millions.

A. Technical Contribution

Our primary contribution is an agent architecture that enables the simultaneous simulation of millions of autonomous agents while maintaining computational efficiency. This represents a fundamental advance in how we model complex social systems - moving from small-scale simulations to population-level modeling that captures emergent behavioral patterns. We've implemented this architecture through AgentTorch, our open-source framework for large-scale agent modeling. While AgentTorch supports our agent architecture, it also provides a broader platform for developing and deploying population-scale AI systems. The framework allows policymakers to test interventions in a simulated environment before real-world implementation, bridging the critical gap between research innovation and practical deployment.

B. The LLM Archetypes Solution

Our research introduces "LLM archetypes," a novel methodology that efficiently integrates LLMs into agent-based models (ABMs) while maintaining the ability to simulate millions of agents. When modeling population-scale phenomena, simulation scale is often more important than individual agent sophistication. The emergent patterns that matter for real-world policy decisions only become visible when simulations operate at true population scale.

LLM archetypes find an optimal balance between behavioral adaptivity and computational efficiency. Rather than choosing between sophisticated but small-scale simulations or simple but large-scale ones, LLM archetypes enable both sophistication and scale simultaneously.

Our experimental results demonstrate that LLM archetypes not only enable simulations to scale to millions of agents, but they also achieve better performance on forecasting and policy evaluation tasks. This performance advantage emerges because archetypes preserve the adaptive, context-aware behaviors that make LLM-guided agents valuable, while capturing the emergent, scale-dependent phenomena that only appear in population-scale simulations.

C. Validation and Real-World Impact

Our research impact spans both experimental validation and real-world policy implementation:

I. Experimental Validation

We conducted experimental validation by creating a digital twin of New York City with 8.4 million autonomous agents to recreate complex patterns of labor force participation and mobility. By validating these simulations against actual census data, we demonstrated two transformative capabilities:

• Efficient Population Monitoring: Large-scale LLM-guided simulations can digitally recreate census-level insights for just a few hundred dollars—presenting an opportunity to move beyond traditional once-in-a-decade census taking toward real-time, passive population monitoring.
• Scale-Sensitive Policy Evaluation: Evaluating policies at true population scale reveals insights that smaller simulations miss, particularly in understanding how behavioral patterns and disease transmission interact during public health crises. These insights simply cannot be captured in small-scale simulations, highlighting why scaling LLM-guided simulations to millions is critical.

II. Real-World Policy Applications

Our research has already translated into real-world impact through our AgentTorch framework. We have partnered with New Zealand's Environmental Science and Research (ESR) institute to create an agent-based digital twin of their society—simulating 5 million citizens and their interactions across health and economic domains.

This system is actively supporting New Zealand's response to the emerging H5N1 bird flu threat, helping authorities understand and manage both public health implications and supply chain disruptions.

D. Technical Implementation: AgentTorch Framework

We've implemented this architecture through AgentTorch, our open-source framework for large-scale agent modeling. While AgentTorch supports our specific LLM archetype architecture, it also provides a broader platform for developing and deploying population-scale AI systems.

The framework allows policymakers to test interventions in a simulated environment before real-world implementation, bridging the critical gap between research innovation and practical deployment.

E. Broader Research Agenda: Large Population Models (LPMs)

This work is part of our broader research agenda on Large Population Models (LPMs), a new paradigm we've proposed for orchestrating beneficial collective behavior at unprecedented scales. LPMs extend beyond traditional agent-based models by incorporating sophisticated coordination protocols that can guide collective behavior toward beneficial outcomes.

F. Looking Forward

Our research represents a significant advancement in how we model complex social systems. By enabling LLM-guided simulations at true population scale, we're opening new possibilities for:

• Creating digital twins of entire societies that can inform policy decisions across domains from public health to economic planning
• Testing counterfactual scenarios that would be impossible to evaluate in the real world
• Capturing emergent phenomena that only appear at population scale

This work demonstrates how academic research in AI can directly address societal challenges, creating a pathway from cutting-edge innovation to meaningful real-world applications.

---

Authors: Ayush Chopra, Shashank Kumar, Nurullah Giray Kuru, Ramesh Raskar, Arnau Quera-bofarull

Key Publications:
- On the limits of agency in agent-based models (Accepted at AAMAS 2025, Oral Presentation)
- The future is now: revolutionising decision-making with AI-driven simulations - AgentTorch in New Zealand