Research

At APES Lab, we develop computational frameworks that enable researchers and policymakers to explore complex socioeconomic systems through multi-agent simulation. Our work bridges theoretical economics with practical policy analysis.

Flagship Framework

SANE v0.5

Simulative Agent-based Normative Environment

SANE is our open-source multi-agent simulation framework designed specifically for policy and economic research. Unlike traditional agent-based models, SANE leverages large language models to create agents with nuanced decision-making capabilities, enabling realistic simulation of human behavior at scale.

  • Scale to millions of heterogeneous agents
  • LLM-powered behavioral modeling for realistic decisions
  • Modular architecture for custom policy environments
  • Validated against historical economic data

SANE v0.5 Architecture

External Inputs
Real-time Market Data
Policy Documents
Historical Records
Survey Data
Policy Environment Engine
LAYER 1
Regulatory Framework
Tax codes, compliance rules, legal constraints
Market Structures
Auction mechanisms, price discovery, liquidity
Institution Rules
Central bank policy, fiscal mechanisms
Environment State
Agent Population Matrix
LAYER 2
👤
Households
4.2B
🏢
Firms
890M
🏦
Banks
45K
🏛️
Gov Entities
12K
LLM Cognitive Core
Memory
Episodic + Semantic
Reasoning
CoT + ToT
Learning
Adaptive + Transfer
Behavioral Signals & Market Actions
Distributed Simulation Engine
LAYER 3
Market Clearing
Real-time price equilibration
📡
Information Propagation
Network diffusion dynamics
📋
Contract Engine
Enforcement & settlement
🔄
State Synchronization
Distributed consensus
Analysis & Projection System
LAYER 4
Aggregate Stats
Macro indicators
Pathway Analysis
Scenario trees
Counterfactuals
What-if modeling
Causal Inference
Effect estimation
Output:
API
Dashboard
Reports
Raw Data

Methodology

Our research methodology combines insights from computational economics, behavioral science, and machine learning to create simulations that capture the complexity of real-world systems.

01

Agent Design

Each agent is endowed with beliefs, preferences, and decision-making capabilities powered by fine-tuned language models. Agents respond to incentives, learn from experience, and adapt to changing environments.

02

Environment Specification

We construct detailed institutional environments that capture market structures, regulatory frameworks, and information flows. These environments are calibrated to match real-world economies.

03

Validation & Analysis

Simulations are validated against historical data and stylized facts. We use ensemble methods and sensitivity analysis to quantify uncertainty and identify robust policy conclusions.

Foundational Research

The academic foundations our work builds upon—pioneering research in agent-based modeling, computational economics, and LLM-driven simulation

Foundational

Growing Artificial Societies: Social Science from the Bottom Up

Epstein, J.M. & Axtell, R.

MIT Press1996

Foundational

Agent-Based Models in Economics: A Toolkit for Policy Analysis

Dawid, H. & Delli Gatti, D.

Handbook of Computational Economics2018

Foundational

The Economy as an Evolving Complex System II

Arthur, W.B., Durlauf, S.N., & Lane, D.A.

Santa Fe Institute Studies1997

Foundational

Macroeconomics from the Bottom-up

Delli Gatti, D., Gaffeo, E., Gallegati, M., et al.

Springer2011

Foundational

Generative Agents: Interactive Simulacra of Human Behavior

Park, J.S., O'Brien, J.C., Cai, C.J., et al.

UIST 20232023

Foundational

Large Language Models as Simulated Economic Agents

Horton, J.J.

NBER Working Paper2023

Foundational
Foundational

SANE v0.5 Specifications

Max Concurrent Agents10,000,000,000
Simulation Time StepsUnlimited
Agent MemoryEpisodic + Semantic
LLM BackendClaude Opus 4.5, GPT 5.1, Gemini 3
Market MechanismsDouble Auction, Posted Price, Search
Policy InstrumentsTax, Subsidy, Regulation, Monetary
Output FormatsJSON, CSV, Parquet, SQL
LicenseApache 2.0