Language-model agents that perceive, deliberate and act reproduce human-like adaptive behavior in epidemic–economic simulations, producing divergent health and economic trajectories and enabling quantitative trade-off analysis; the framework is robust across several LLMs but requires empirical calibration before real-world policy use.

Main Finding

LLM-driven agents embedded in a Perception–Deliberation–Action (PDA) loop produce endogenous, human-like adaptive behaviors via Chain-of-Thought reasoning. When used in a coupled epidemic–economic simulation, this framework robustly generates divergent macro trajectories and quantitatively captures the trade-offs between public health outcomes and economic stability across a range of macroscopic scenarios and LLM backends — effectively creating a high-fidelity computational laboratory that links micro-level cognition to macro-level societal outcomes.

Key Points

• Framework innovation: Replaces static, rule-based agent decision-making with LLM-powered cognitive agents that perceive environment signals, deliberate using Chain-of-Thought, and act—without hand-coded behavior rules.
• Perception–Deliberation–Action (PDA) loop:
  • Perception: agents observe local epidemic and economic signals (e.g., infection risk, income/work opportunities).
  • Deliberation: agents use LLM Chain-of-Thought to reason about trade-offs (health risk vs. economic necessity).
  • Action: agents select behaviors (e.g., alter social contacts, participate in work, change consumption) that feed back into epidemic and economic dynamics.
• Endogenous adaptation: Behavioral changes emerge from cognitive reasoning rather than parameterized switches, producing context-sensitive, heterogeneous responses.
• Robustness: Experiments run with multiple LLM backends (proprietary and open-source) show qualitatively consistent dynamics, indicating framework stability to model choice.
• Outcomes: The model produces divergent socio-economic trajectories under different macroscopic conditions and quantitatively characterizes the public-health vs. economic-stability trade-off (e.g., differing infection curves and economic activity levels).
• Usefulness: Enables scenario testing for policies and shocks where human judgment and adaptation matter (lockdowns, targeted interventions, information campaigns).

Data & Methods

• Model architecture:
  • Multi-agent simulation coupling an epidemic module (transmission depends on agent interactions and behaviors) with an economic module (aggregate activity driven by agent-level labor/consumption decisions).
  • Each agent implemented as an LLM-driven cognitive unit running the PDA loop each timestep.
  • Chain-of-Thought prompts/internal reasoning are used to simulate richer, multi-step decision processes.
• Heterogeneity:
  • Agents differ in their perceptions, priorities (health vs. income), and local conditions, producing diverse responses.
• LLM backends:
  • Framework evaluated across several LLMs (multiple proprietary and open-source models) to test sensitivity to language model choice.
• Experiments:
  • Diverse macroscopic scenarios (e.g., varying baseline transmissibility, economic shocks, information environments, policy regimes).
  • Metrics tracked at micro and macro scales: individual behavior patterns, infection prevalence over time, and aggregate economic indicators (changes in activity/employment/consumption).
  • Quantitative analysis of trade-offs between health metrics and economic outcomes across scenarios and model variants.
• Validation & robustness checks:
  • Cross-backend comparisons and scenario sweeps to assess qualitative consistency and sensitivity.
  • (Paper notes necessity of further empirical calibration; framework primarily demonstrates method and emergent phenomena.)

Implications for AI Economics

• Better behavioral realism: LLM-driven agents can produce richer, context-sensitive adaptive behavior, improving the realism of economic-epidemic simulations where human judgment matters.
• Policy evaluation: The framework enables testing of nuanced interventions (timing, targeting, communication strategies) and forecasts their macro impacts considering endogenous behavioral change.
• Micro–macro linkage: Provides a tractable way to endogenize micro-level cognition and see how cognitive heterogeneity aggregates into systemic outcomes and trade-offs.
• Scenario analysis and risk assessment: Useful for exploring tail events, cascading failures, and distributional consequences when behavior is a key driver.
• Research directions:
  • Empirical calibration and validation against observed behavioral/economic data to increase predictive credibility.
  • Incorporating LLM alignment, bounded rationality constraints, and memory or learning across episodes to better match human cognition.
  • Cost/compute and reproducibility improvements (e.g., distilled/smaller LLMs, surrogate models for large runs).
• Cautions:
  • Dependence on LLM behavior: hallucinations, bias, or misaligned reasoning in language models can propagate into simulated outcomes.
  • Interpretability and auditability: Chain-of-Thought reasoning may be hard to fully verify; careful logging and post-hoc analysis required.
  • Ethical and policy risks: Simulations could be misused without proper transparency; real-world policy recommendations require robust validation.

Overall, LLM-driven multi-agent simulations offer a promising step toward endogenizing human-like adaptive decision-making in economic models, enabling more realistic analysis of policy trade-offs during crises — while necessitating careful validation, cost-management, and safeguards.