Awesome Learned Social Simulation Engines

A curated, functional map of systems that learn, perturb, and validate the dynamics of bounded social systems.

Learned social simulation engines are composable, perturbable models of complex human systems. Unlike hand-coded simulations, they learn dynamics from data, represent uncertainty as probability landscapes rather than point predictions, and support rigorous intervention testing in silico. This list curates the methods, tools, datasets, and evaluation practices for building and understanding them — each entry annotated by what it contributes to an engine, not by topic alone.

Who this is for

• Computational social scientists and agent-based / multi-agent modellers
• AI engineers building agent societies and synthetic populations
• Policy and intervention modellers
• Research agents needing a curated, machine-readable substrate

How to use this list

• Browse by domain in the Resource Map — labour markets, mobility, information diffusion, institutions, and more.
• Read across by function in the Functional lens — reconstruct, simulate, plan, calibrate, validate, risk.
• Every entry carries a type label and a one-line statement of its contribution.

What belongs

• Primary research, maintained tools, canonical datasets, and benchmarks that advance the learned simulation of social systems.

What does not

• Generic AI/ML material with no social-simulation contribution, promotional pages, or unmaintained tools.

The engine vision

Rather than a single monolithic simulator, the field is moving toward a modular landscape of tool models, each representing a bounded slice of social complexity — labour markets, urban mobility, information diffusion, institutional behaviour — composed through a shared interface and read out as probability landscapes rather than point forecasts.

flowchart TD
    Data["Empirical data and traces"]
    Policy["Policy interventions"]

    Data --> API
    Policy --> API

    subgraph Engine["Learned Social Simulation Engine"]
        API["API-driven composability layer"]

        subgraph Slices["Bounded social-system slices"]
            Labour["Labour markets"]
            Urban["Urban mobility"]
            Info["Information diffusion"]
            Institution["Institutional behaviour"]
        end

        API --> Labour
        API --> Urban
        API --> Info
        API --> Institution
    end

    Labour --> Dynamics
    Urban --> Dynamics
    Info --> Risks
    Institution --> Risks

    Dynamics["Emergent social dynamics"]
    Risks["Distributional risks and harms"]

    Dynamics --> Landscapes
    Risks --> Landscapes

    Landscapes["Probability landscapes"]

Resource Map

The list is grouped into tiers for navigation; the sections themselves are the primary taxonomy.

Foundations

Methods

Engines, systems & composability

Reasoning & intervention

Calibration, validation & benchmarks

Risk & responsibility

Substrate & tooling

Parallels & documentation

Entry types: paper peer-reviewed paper or preprint · book · article essay, report, or news feature · tool library, framework, or platform · framework standard, protocol, or reference framework · dataset · chapter book chapter.

Functional lens

The domain sections are the primary taxonomy. This functional lens is a secondary, cross-cutting view: it groups systems by what they do inside an engine, adapting the renderer / simulator / planner distinction from world-model research to social systems.

• Reconstruct — produce or reconstruct observable social state: synthetic populations, digital twins, generative agents as a population substrate.
• Simulate — model transitions, interactions, and counterfactuals: agent-based models, microsimulation, multi-agent reinforcement learning, learned world models, neural surrogates.
• Plan — use simulated futures to choose actions: reinforcement learning over simulators, policy and mechanism design, causal intervention layers.

Social systems lack the invariant laws of physical world models, so three further cross-cutting categories are first-class rather than optional:

• Calibrate — fit a simulator to data: simulation-based inference, history matching.
• Validate — establish trust: uncertainty quantification, behavioural-fidelity benchmarks, model documentation.
• Risk — surface second-order and distributional harms, and leakage from synthetic populations.

The lens is piloted on Existing Systems below; tags can be extended to other sections over time.

Resources

Computational Social Science

• Lazer et al. — Computational Social Science: Obstacles and Opportunities (2020) paper — A decade-on state-of-the-field assessment; identifies what large-scale simulation, observational data, and experimental design can now achieve and what structural barriers remain for policy-relevant CSS.
• Gonzalez-Bailon et al. — Computational Social Science and Sociology (2020) paper — Maps how computational methods — including simulation — are reshaping sociology; a practical orientation to the field for researchers building social engines.
• Bail, C. — Breaking the Social Media Prism (2021) book — Uses large-scale social simulation experiments to test theories of political polarisation; a worked example of CSS methods applied to a real policy-relevant question.

Complex Systems

• Mitchell, M. — Complexity: A Guided Tour (2009) book — The clearest single-volume primer on emergence, feedback loops, adaptation, and self-organisation; the conceptual foundation for understanding why social simulation engines behave the way they do.
• Aeon — Complex systems science allows us to see new paths forward (2024) article — Connects complex systems thinking to contemporary social challenges including pandemics and inequality; explains why non-linearity and emergence make single-point predictions insufficient.
• Farmer & Foley — The economy needs agent-based modelling (Nature 2009) article — Argues that agent-based simulation engines are the right tool for economics precisely because social systems are complex and adaptive; a direct intellectual ancestor of the simulation engine concept.

Social Simulation

• Park et al. — Generative Agents: Interactive Simulacra of Human Behavior (2023) paper — The landmark paper demonstrating LLM-driven agents that wake up, form relationships, and coordinate events in a simulated town; the most-cited recent example of a working social simulation engine.
• Gao et al. — AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents (2025) paper — Proposes a large-scale social simulator with a dedicated simulation engine, realistic societal environment, and 10,000+ LLM-driven agents; directly demonstrates the architecture this repo maps toward.
• Nature — The first 'AI societies' are taking shape (2025) article — A Nature news feature surveying the emerging landscape of AI-powered social simulation engines and the scientific and ethical questions they raise.

Microsimulation

• Katikireddi et al. — Microsimulation as a flexible tool to evaluate policies and their impact on socioeconomic inequalities in health (2023) paper — Demonstrates how a microsimulation engine surfaces unequal health effects of policy interventions across income and demographic groups — a direct example of the distributional analysis this engine aims to support.
• Lovelace et al. — A Large-Scale Geographically Explicit Synthetic Population with Social Networks for the United States (2024) paper — Builds a country-scale synthetic population with embedded social networks for agent-based and microsimulation use; shows the current state of the art in population engine construction.

Causal Inference

• Hernán, M.A. & Robins, J.M. — Causal Inference: What If (2023) book — The definitive free textbook on causal inference using potential outcomes; covers counterfactual reasoning and intervention analysis directly applicable to simulation engine design. Freely available online.
• Kıcıman et al. — Causal Reasoning and Large Language Models: Opening a New Frontier for Causality (2023) paper — Evaluates whether LLMs can support causal reasoning in simulation contexts; directly relevant to building intervention layers in LLM-augmented social engines.
• Nilforoshan et al. — Language Models as Scientist and Subjects (2024) paper — Uses structural causal models combined with LLMs to automatically generate and test social scientific hypotheses in silico; a working example of causal simulation engine methodology.
• Imbens, G.W. & Rubin, D.B. — Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction (2015) book — Stanford / Harvard reference for the Rubin potential-outcomes framework — randomised experiments, observational studies, propensity scores, instrumental variables; the standard companion to the SCM tradition for any simulation engine reasoning about interventions.
• Pearl, J. — Causality: Models, Reasoning, and Inference, 2nd ed. (2009) book — UCLA monograph that establishes structural causal models, do-calculus, and the counterfactual hierarchy as the mathematical foundation for intervention reasoning — the SCM bedrock the rest of this section sits on.
• Sharma, A. & Kıcıman, E. — DoWhy: An End-to-End Library for Causal Inference (2020) tool — Microsoft Research Python library that unifies graphical-model and potential-outcomes workflows behind a model–identify–estimate–refute pipeline; the most opinionated Python entry point for principled causal estimation with explicit assumption-testing.
• Battocchi et al. — EconML: A Python Package for ML-Based Heterogeneous Treatment Effects Estimation (2019) tool — Microsoft Research ALICE project library implementing double machine learning, doubly robust learners, causal forests, and meta-learners; the canonical Python tool for heterogeneous treatment effect estimation at scale.
• Chen, Harinen, Lee, Yung & Zhao — CausalML: Python Package for Causal Machine Learning (2020) tool — Uber Engineering library focused on uplift modelling and CATE estimation via meta-learners (S/T/X/R), uplift trees, and causal forests; widely used in industrial experimentation and policy targeting.
• Ankan, A. & Textor, J. — pgmpy: A Python Toolkit for Bayesian Networks (JMLR 2024) tool — Maintained Python toolkit covering structure learning, parameter estimation, exact and approximate inference, and DAG-based causal reasoning; the practical complement to DoWhy/EconML when the modelling target is the structure itself, not just the effect.
• Zheng, Aragam, Ravikumar & Xing — DAGs with NO TEARS: Continuous Optimization for Structure Learning (NeurIPS 2018) paper — CMU reformulation of DAG structure learning as a smooth continuous optimisation problem; foundational paper for AI-native causal discovery and the seed of the gradient-based causal-discovery literature that followed.
• Xia, Lee, Bengio & Bareinboim — The Causal-Neural Connection: Expressiveness, Learnability, and Inference (NeurIPS 2021) paper — Columbia / Mila result formalising Neural Causal Models (NCMs) and showing that universal-approximation neural nets do not automatically learn the right SCM; the principled basis for combining deep learning with Pearl-style causal inference.
• Dyer, Bishop, Felekis, Zennaro, Calinescu, Damoulas & Wooldridge — Interventionally Consistent Surrogates for Complex Simulation Models (NeurIPS 2024) paper — Oxford INET / CS application of causal abstraction to large agent-based and complex simulators; learns surrogate models guaranteed to behave consistently under interventions of interest, the missing piece for fast counterfactual policy testing on heavy simulation engines.

Uncertainty Quantification

• Kahn et al. — Uncertainty Quantification for Agent-Based Models (2024) paper — Applies Gaussian processes, sequential design, and history matching to quantify uncertainty in ABM outputs; the current state of the art for making social simulation engines produce calibrated rather than overconfident results.
• Salter et al. — Improving Policy-Oriented Agent-Based Modeling with History Matching (2025) paper — Demonstrates history matching as a principled method for calibrating policy-oriented ABMs against real-world data; directly applicable to validating social simulation engines.
• Saltelli et al. — Global Sensitivity Analysis: The Primer (2008) book — The standard reference for sensitivity analysis of model outputs; explains how to identify which uncertain inputs drive output variance — still the foundational text for this capability.
• Cranmer, Brehmer & Louppe — The Frontier of Simulation-Based Inference (PNAS 2020) paper — NYU / Liège review naming and framing neural simulation-based inference as a research field; the canonical entry-point that places NPE / NLE / NRE alongside classical ABC and explains why amortized neural posteriors are the right tool when a simulator's likelihood is intractable.
• Papamakarios, Sterratt & Murray — Sequential Neural Likelihood: Fast Likelihood-free Inference with Autoregressive Flows (AISTATS 2019) paper — Edinburgh paper training autoregressive-flow likelihood emulators on simulator runs; one of the foundational neural-likelihood-estimation methods that turn an opaque ABM or mechanistic simulator into a usable Bayesian model.
• Hermans, Delaunoy, Rozet, Wehenkel, Begy & Louppe — A Trust Crisis in Simulation-Based Inference? Your Posterior Approximations Can Be Unfaithful (TMLR 2022) paper — Liège-led empirical demonstration that neural SBI posteriors can be silently overconfident; the diagnostic warning every team adopting SBI for a social simulator should read before trusting a "calibrated" posterior.
• Tejero-Cantero, Boelts, Deistler, Lueckmann, Durkan, Gonçalves, Greenberg & Macke — sbi: A toolkit for simulation-based inference (JOSS 2020) tool — Macke lab (Tübingen) open-source Python toolkit implementing NPE, NLE, NRE, and their sequential variants behind a uniform API; the practical tool researchers reach for to put neural SBI on a real simulator.
• Lueckmann, Boelts, Greenberg, Gonçalves & Macke — Benchmarking Simulation-Based Inference (AISTATS 2021) paper — Macke lab benchmark suite covering ten simulator tasks against NPE / NLE / NRE / SMC-ABC; the reference comparison that lets a practitioner choose an SBI algorithm for a given problem rather than picking by reputation.
• Dyer, Cannon, Farmer & Schmon — Black-box Bayesian inference for economic agent-based models (2022) paper — Oxford INET application of neural simulation-based inference to large economic ABMs with intractable likelihoods; a working template for amortized posterior calibration of social ABMs and the methodological complement to history-matching approaches above.
• Andrianakis, Vernon, McCreesh, McKinley, Oakley, Nsubuga, Goldstein & White — Bayesian History Matching of Complex Infectious Disease Models Using Emulation (PLOS Comp Bio 2015) paper — Cambridge MRC / Durham tutorial-and-case-study fitting an HIV-Uganda transmission simulator via Gaussian-process emulators and waves of implausibility cuts; the foundational Bayesian-emulation methodology that Kahn 2024 and Salter 2025 build on for ABMs.
• Kasim et al. — Building high accuracy emulators for scientific simulations with deep neural architecture search (MLST 2021) paper — Oxford-led demonstration that deep-neural-network emulators discovered by automated architecture search achieve mega-fold speedups across ten scientific simulators (climate, fusion, astrophysics); the AI-native upgrade path from GP-only emulators to deep emulators for high-dimensional social simulators.
• Talts, Betancourt, Simpson, Vehtari & Gelman — Validating Bayesian Inference Algorithms with Simulation-Based Calibration (2018) paper — Columbia / Aalto methodology for checking whether a Bayesian inference algorithm is itself well-calibrated by drawing from the prior, simulating data, and verifying rank-uniformity of true-vs-estimated posteriors; the standard diagnostic any neural-SBI pipeline applied to a social simulator should pass.

Responsible AI

• Weidinger et al. — Sociotechnical Safety Evaluation of Generative AI Systems (2023) paper — Proposes a framework for evaluating harms from generative AI systems across social, institutional, and individual dimensions; directly applicable to responsible design of LLM-driven simulation engines.
• Obermeyer et al. — Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations (2019) paper — The canonical empirical demonstration that an AI system producing good aggregate outcomes can simultaneously produce severe distributional harm; the case study every simulation engine designer should know.
• NIST — Artificial Intelligence Risk Management Framework (AI RMF 1.0) (2023) framework — The leading voluntary framework for managing AI risk across the full system lifecycle; covers trustworthiness, accountability, bias, and governance for AI-assisted policy and simulation tools.

Existing Systems

Functional lens pilot — primary role of each system (see Functional lens):

• Vezhnevets et al. — Concordia: Generative Agent-Based Modelling with Actions Grounded in Physical, Social, or Digital Space (2023) tool — Google DeepMind's open-source library for building LLM-driven multi-agent simulations grounded in physical or digital environments; the reference engine from a leading AI lab for social science modelling with generative agents.
• Altera.AL — Project Sid: Many-Agent Simulations Toward AI Civilization (2024) paper — Demonstrates 10–1,000+ AI agents developing specialised professions, governance structures, cultural norms, and economic systems in a shared environment; the most ambitious published test of emergent civilisational dynamics in an AI simulation engine.
• Yang et al. — OASIS: Open Agent Social Interaction Simulations with One Million Agents (2024) tool — Scalable social media simulator supporting up to one million LLM-driven agents with dynamic social networks, recommendation systems, and diverse action spaces; designed to study large-scale group behaviour and information propagation.
• Argyle et al. — Out of One, Many: Using Language Models to Simulate Human Samples (2023) paper — Stanford-affiliated work demonstrating that LLMs conditioned on demographic profiles can reproduce the distribution of human survey responses; foundational evidence for using AI agents as synthetic populations in social simulation.
• Stanford Digital Economy Lab — Economic Simulations with AI tool — Stanford research programme combining LLMs, agent-based methods, and causal reasoning to build transparent, auditable economic simulation environments; an active lab-scale effort to build AI-native policy simulation engines.

Agent-Based Modelling

• Kazil, Masad & Crooks — Utilizing Python for Agent-Based Modeling: The Mesa Framework (2020) tool — Mesa is the canonical Python ABM framework: built-in spatial and network grids, browser-based visualisation, and seamless integration with the scientific Python stack — the default starting point for researchers building ABMs in Python.
• Foramitti — AgentPy: A package for agent-based modeling in Python (2021) tool — Pythonic ABM framework with first-class support for sensitivity analysis, parameter exploration, and Jupyter-native experiment workflows; designed for the experimental-design discipline a simulation engine needs.
• Wilensky — NetLogo (1999) tool — The most widely taught ABM environment: low floor for prototyping, large library of curated example models, and the de-facto reference implementation for emergence pedagogy and exploratory modelling.
• Collier & Ozik — Distributed Agent-Based Simulation with Repast4Py (2022) tool — Argonne's MPI-distributed Python ABM toolkit; the practical path to running models that don't fit on one machine without abandoning Python.
• Richmond et al. — FLAME GPU 2 (2023) tool — University of Sheffield framework for GPU-accelerated ABM at population scale; brings million-agent simulations into reach on commodity GPU hardware.
• Datseris, Vahdati & DuBois — Agents.jl: a performant and feature-full agent-based modeling software (2022) tool — Julia ABM framework benchmarking faster and lower-code than mainstream alternatives; the option when ABM performance, not just convenience, is the constraint.
• Grimm et al. — The ODD Protocol for Describing Agent-Based and Other Simulation Models: A Second Update (2020) framework — The standard documentation protocol for ABMs; without ODD-style description, simulation engines fail the basic replicability bar this list cares about.
• Epstein — Generative Social Science: Studies in Agent-Based Computational Modeling (2006) book — The foundational manifesto for ABM as theory-building: "if you didn't grow it, you didn't explain it" — the methodological commitment that distinguishes simulation engines from predictive models.
• Railsback & Grimm — Agent-Based and Individual-Based Modeling: A Practical Introduction, 2nd ed. (2019) book — The standard textbook on ABM design, programming, documentation, and analysis; the closest thing the field has to a shared training curriculum.
• Schelling — Dynamic Models of Segregation (1971) paper — The canonical emergence demonstration: simple individual rules generating large-scale segregation that no agent intends; still the clearest worked example of why aggregate outcomes mislead and bottom-up modelling is needed.
• Epstein & Axtell — Growing Artificial Societies: Social Science from the Bottom Up (1996) book — The Sugarscape monograph: artificial societies in which trade, wealth distributions, group conflict, and cultural transmission emerge from minimal local rules; the prototype for what a social-science simulation engine grows.
• Chopra et al. — On the Limits of Agency in Agent-Based Models (AAMAS 2025) paper — MIT Media Lab work introducing LLM-driven behaviour archetypes inside a differentiable ABM (AgentTorch), scaling LLM-guided agent simulations to millions; a methodological contribution distinct from generative-agent work in ### Social Simulation.
• Zheng et al. — The AI Economist: Taxation policy design via two-level deep multiagent reinforcement learning (2022) paper — Salesforce Research two-level RL framework where economic agents and a social planner co-adapt; a worked example of fusing multi-agent RL with ABM-style economic simulation for policy design.

Synthetic Populations

• Borysov, Rich & Pereira — How to generate micro-agents? A deep generative modeling approach to population synthesis (2019) paper — Technical University of Denmark; the canonical VAE-for-population-synthesis paper, framing why deep generative methods can recover sampling-zero attribute combinations that IPF/IPU systematically miss.
• Garrido, Borysov, Pereira & Rich — Composite Travel Generative Adversarial Networks for Tabular and Sequential Population Synthesis (2020) paper — DTU follow-up extending population synthesis to GANs that jointly generate tabular agent attributes and sequential mobility traces; shows how generative models can produce structurally consistent micro-agents, not just marginally correct ones.
• Tang, Lu & Feng — Generating Feasible and Diverse Synthetic Populations Using Diffusion Models (2025) paper — Diffusion-based population synthesis that improves the feasibility/diversity trade-off over VAEs and GANs; the current methodological frontier for generative population modelling.
• Cao, Liu, Arora, Augenstein, Röttger & Hershcovich — Specializing Large Language Models to Simulate Survey Response Distributions for Global Populations (NAACL 2025) paper — University of Copenhagen / Bocconi work fine-tuning LLMs to match survey response distributions across global populations; treats LLM-as-population as a calibration target rather than a black-box prompt.
• Aher, Arriaga & Kalai — Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies (ICML 2023) paper — Microsoft Research framework for "Turing Experiments" where an LLM is used to simulate a representative sample of participants in classic behavioural studies; the methodological precursor to LLM-as-population research.
• Horton, Filippas & Manning — Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus? (NBER 2023) paper — MIT/Fordham argument that LLMs are implicit computational models of humans ("Homo silicus") that economists can endow, prompt, and interrogate as a population sample; the conceptual foundation for LLM-conditioned synthetic populations.
• Törnberg, Valeeva, Uitermark & Bail — Simulating Social Media Using Large Language Models to Evaluate Alternative News Feed Algorithms (2023) paper — Amsterdam + Duke construction of an LLM persona population grounded in American National Election Studies microdata; a worked example of using real survey microdata to anchor LLM-driven synthetic societies.
• Tencent AI Lab — Scaling Synthetic Data Creation with 1,000,000,000 Personas (2024) paper — Tencent AI Lab persona-driven data synthesis at billion-persona scale; demonstrates how LLM-derived personas can serve as a population substrate for diverse downstream synthesis tasks.
• Park et al. — Generative Agent Simulations of 1,000 People (2024) paper — Stanford + Google DeepMind construction of generative agents from two-hour interviews with 1,000 demographically representative US participants; agents replicate participants' General Social Survey responses at 85% of test–retest reliability.
• Toubia, Gui, Peng, Merlau, Li & Chen — Twin-2K-500: A Data Set for Building Digital Twins of over 2,000 People Based on Their Answers to over 500 Questions (Marketing Science 2025) tool — Columbia Business School digital-twin dataset of 2,058 US participants across demographic, psychological, economic, and behavioural-experiment measures; an open substrate for training and evaluating LLM-based individual digital twins.
• Bilal et al. — CitySEIRCast: An Agent-Based City Digital Twin for Pandemic Analysis and Simulation (Complex & Intelligent Systems 2024) paper — Notre Dame / Brunel city-scale digital-twin construction coupling synthetic urban populations with mobility, social-media, and contact data; the practical model for sub-national AI-native population builds for epidemic analysis.
• Kashiyama, Pang, Sekimoto & Yabe — Pseudo-PFLOW: Nationwide Synthetic Human Mobility Dataset Construction from Limited Travel Surveys and Open Data (Computer-Aided Civil and Infrastructure Engineering 2024) paper — Tokyo / Osaka research-lab synthetic population covering all ~130 million people across Japan's 47 prefectures; a non-Western, country-scale AI-native population build that closes a known coverage gap.
• Wang et al. — Population Synthesis with Deep Generative Model: A Joint Household-Individual Approach (Computational Urban Science 2025) paper — Joint household-and-individual deep generative framework; addresses the structural-zero problem that pure individual-level synthesis can't solve, important for downstream agent-based and microsimulation use.

Cross-references: Lovelace et al. 2024 US synthetic population is listed under ### Microsimulation; Argyle et al. 2023 LLM-as-sample work is under ### Existing Systems; Park et al. 2023 Generative Agents is under ### Social Simulation; Chopra et al. 2025 AgentTorch is under ### Agent-Based Modelling. They are deliberately not duplicated here.

Multi-Agent Reinforcement Learning

• Terry et al. — PettingZoo: Gym for Multi-Agent Reinforcement Learning (NeurIPS 2021) tool — Farama Foundation Gym-style API standardising how multi-agent environments expose observations, actions, rewards, and turn order; the de-facto entry-point library for AI-native MARL research and the substrate most other tools in this section integrate with.
• Lanctot et al. — OpenSpiel: A Framework for Reinforcement Learning in Games (DeepMind 2019) tool — DeepMind framework hosting hundreds of games (cooperative, competitive, perfect / imperfect information) with both classical game-theoretic algorithms and deep-RL training; the canonical multi-game testbed when the social-science question crosses normal-form, extensive-form, and repeated games.
• Suarez, Du, Isola & Mordatch — Neural MMO: A Massively Multiagent Game Environment for Training and Evaluating Intelligent Agents (2019) tool — MIT massively-multi-agent persistent-world environment inspired by MMORPGs supporting hundreds of concurrent learning agents; the reference open-ended environment for studying emergent specialisation, division of labour, and population-scale dynamics.
• Liang et al. — RLlib: Abstractions for Distributed Reinforcement Learning (ICML 2018) tool — UC Berkeley / Anyscale distributed-RL library with first-class multi-agent abstractions and PettingZoo compatibility; the practical training infrastructure when MARL experiments outgrow a single machine.
• Leibo, Zambaldi, Lanctot, Marecki & Graepel — Multi-agent Reinforcement Learning in Sequential Social Dilemmas (AAMAS 2017) paper — DeepMind paper introducing sequential social dilemmas — Markov games (Gathering, Wolfpack) where collective rationality is at odds with individual gradient descent; the foundational benchmark family for studying cooperation, defection, and free-riding under deep learning.
• Agapiou et al. — Melting Pot 2.0 (DeepMind 2023) tool — DeepMind evaluation suite of 50+ scenarios specifically built to test how MARL agents generalise to novel social situations and partner populations; the most thorough behavioural testbed for cooperation, exploitation, and norm-following in deep MARL.
• Carroll, Shah, Ho, Griffiths, Seshia, Abbeel & Dragan — On the Utility of Learning about Humans for Human-AI Coordination (NeurIPS 2019) paper — UC Berkeley Overcooked-based environment testing whether RL agents trained via self-play actually coordinate with humans; the canonical demonstration that pure self-play is insufficient for human-AI coordination and the trigger for the human-aware-coordination literature.
• Lowe, Wu, Tamar, Harb, Abbeel & Mordatch — Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments (MADDPG, NeurIPS 2017) paper — OpenAI / UC Berkeley centralised-critic actor-critic where each agent learns a policy with access to all agents' observations during training; the seminal centralised-training-decentralised-execution method for mixed cooperative-competitive deep MARL.
• Rashid, Samvelyan, Schroeder de Witt, Farquhar, Foerster & Whiteson — QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning (ICML 2018) paper — Oxford WhiRL value-decomposition method factorising a joint Q-function as a monotonic mixture of per-agent Q-functions; the canonical baseline for cooperative MARL with global rewards under partial observability.
• Foerster, Farquhar, Afouras, Nardelli & Whiteson — Counterfactual Multi-Agent Policy Gradients (COMA, AAAI 2018) paper — Oxford counterfactual baseline for centralised actor-critic MARL — marginalises out one agent's action to isolate its contribution to the team reward; foundational credit-assignment method for fully cooperative deep MARL.
• Yu, Velu, Vinitsky, Gao, Wang & Bayen — The Surprising Effectiveness of PPO in Cooperative Multi-Agent Games (MAPPO, NeurIPS 2022 D&B) paper — Tsinghua / MIT / UC Berkeley demonstration that on-policy PPO with shared parameters and centralised value functions matches or beats specialised cooperative MARL methods on standard benchmarks; the strong baseline that re-anchored the algorithmic state of the art.
• Meta FAIR Diplomacy Team — Human-level play in the game of Diplomacy by combining language models with strategic reasoning (CICERO, Science 2022) paper — Meta FAIR agent achieving human-level no-press Diplomacy play by combining a dialogue language model with a self-play-trained strategic reasoning module; the landmark demonstration that LLM-MARL hybrids can coordinate over natural language in social games at human level.
• Strouse, McKee, Botvinick, Hughes & Everett — Collaborating with Humans without Human Data (Fictitious Co-Play, NeurIPS 2021) paper — DeepMind fictitious-co-play method training a single agent against a population of self-play partners and then deploying it with humans; the practical recipe for MARL agents that coordinate with novel human partners zero-shot.

Cross-references: Vezhnevets et al. 2023 Concordia and Altera.AL 2024 Project Sid are listed under ### Existing Systems; Park et al. 2023 Generative Agents is under ### Social Simulation; Chopra et al. 2025 AgentTorch and Zheng et al. 2022 The AI Economist are under ### Agent-Based Modelling. They touch MARL but are deliberately not duplicated here.

LLM-Based Social Simulation

• Bougie, Ye & Watanabe — CityReal: Human-Aligned Urban Behavior and City Dynamics Simulation with Large-Scale LLM Agents (2026) paper — Large-scale human-aligned urban simulation where intention-driven LLM agents learn habits and preferences via textual adapters to match real population statistics at micro and macro scale.
• Park, Popowski, Cai, Morris, Liang & Bernstein — Social Simulacra: Creating Populated Prototypes for Social Computing Systems (UIST 2022) paper — Stanford predecessor to Generative Agents using GPT-3 to populate a Reddit-style community for design-research prototyping; the foundational demonstration that an LLM can simulate a plausible online community at sufficient fidelity for system designers to act on its outputs.
• Packer, Wooders, Lin, Fang, Patil, Stoica & Gonzalez — MemGPT: Towards LLMs as Operating Systems (2023) paper — UC Berkeley Sky Lab virtual-context architecture that pages information in and out of an LLM's working window using OS-style memory hierarchies; the canonical persistent-memory contribution for long-running social agents that need horizon-spanning coherence.
• Wang, Xie, Jiang, Mandlekar, Xiao, Zhu, Fan & Anandkumar — Voyager: An Open-Ended Embodied Agent with Large Language Models (NeurIPS 2023) paper — NVIDIA / Caltech lifelong LLM agent with an automatic curriculum, growing skill library, and iterative self-verification loop; the reference architecture for open-ended autonomous behaviour reused across generative-agent designs.
• Shinn, Cassano, Berman, Gopinath, Narasimhan & Yao — Reflexion: Language Agents with Verbal Reinforcement Learning (NeurIPS 2023) paper — Northeastern / Princeton method using verbal self-reflection on prior trajectories as a learning signal in place of weight updates; the canonical reflection mechanism cited across generative-agent architectures including Park et al. 2023.
• Li, Hammoud, Itani, Khizbullin & Ghanem — CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society (NeurIPS 2023) tool — KAUST role-playing framework that pairs LLM agents under "inception prompting" so they autonomously cooperate on a task with minimal human-written instructions; the foundational LLM-society scaffolding paper and a widely-used substrate for studying agent cooperation.
• Wu, Bansal, Zhang, Wu, Li, Zhu, Jin, Liu, Yang, Chen, Wang & Wang — AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation (2023) tool — Microsoft Research open-source framework defining customisable conversational LLM agents with tool use, code execution, and human-in-the-loop hooks; the de-facto general-purpose LLM-multi-agent toolkit researchers reach for to compose agent populations.
• Hong, Zheng, Chen, Cheng, Wang, Zhang, Wang, Yau, Lin, Zhou, Ran, Xiao, Wu & Schmidhuber — MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework (ICLR 2024 oral) tool — DeepWisdom framework encoding human Standard Operating Procedures as structured message-passing protocols between role-specialised LLM agents; the most-cited methodology paper on structured LLM-multi-agent coordination.
• Qian, Liu, Liu, Chen, Dang, Li, Yang, Chen, Su, Cong, Xu, Li, Liu & Sun — ChatDev: Communicative Agents for Software Development (ACL 2024) tool — Tsinghua / OpenBMB virtual-company simulation with role-specialised LLM agents (CEO, CTO, programmer, tester) collaborating through chat; a worked LLM-society case study and a widely-reused multi-agent scaffold for organisational simulation.
• Chen, Su, Zuo, Yang, Yuan, Qian, Chan, Qin, Lu, Xie, Liu, Sun & Zhou — AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors (ICLR 2024) tool — Tsinghua framework explicitly built to study how groups of LLM agents exhibit emergent behaviours under collaborative tasks; one of the few multi-agent toolkits whose stated aim is social-emergence research rather than task throughput.
• Liu, Yang, Jia, Zhang, Yang & Vosoughi — Training Socially Aligned Language Models on Simulated Social Interactions (Stable Alignment, ICLR 2024) paper — Dartmouth / Stanford alignment method that trains an LLM by exposing it to a sandboxed society of LLM agents giving each other social feedback; the canonical demonstration that an LLM can be socially aligned through in-silico simulated interaction rather than direct human ratings.
• Santurkar, Durmus, Ladhak, Lee, Liang & Hashimoto — Whose Opinions Do Language Models Reflect? (ICML 2023) paper — Stanford evaluation of how LLM opinions on US public-policy issues align with the responses of different demographic groups; foundational evidence that LLM-as-population work inherits a measurable, non-uniform population skew any norm-formation simulation must account for.
• Chuang, Goyal, Harlalka, Suresh, Hawkins, Yang, Shah, Hu & Rogers — Simulating Opinion Dynamics with Networks of LLM-based Agents (NAACL 2024 Findings) paper — University of Wisconsin–Madison study running canonical opinion-dynamics models with LLM agents on social-network topologies; an empirical anchor for whether LLM populations reproduce known sociological dynamics like polarisation and consensus formation.
• CoevolveSim — Belief Coevolution in a Social Network of Generalist and Specialist Large Language Models (2026) paper — Introduces a networked LLM simulation for studying how beliefs coevolve when generalist and specialist agents interact.
• Mei, Xie, Yuan & Jackson — A Turing test of whether AI chatbots are behaviorally similar to humans (PNAS 2024) paper — Michigan / Stanford head-to-head behavioural comparison of leading LLMs against humans across canonical economic games and Big-Five personality measures; the cleanest-cited Turing-style empirical anchor for whether LLM agents behave like humans in social tasks.
• Salvi, Horta Ribeiro, Gallotti & West — On the Conversational Persuasiveness of Large Language Models: A Randomized Controlled Trial (Nature Human Behaviour 2024) paper — EPFL pre-registered RCT showing that GPT-4 with light personalisation out-persuades humans in head-to-head debates; landmark empirical result on LLM social influence and a methodological template for measuring persuasion effects in simulated societies.

Cross-references: Vezhnevets et al. 2023 Concordia, Altera.AL 2024 Project Sid, Yang et al. 2024 OASIS and Argyle et al. 2023 Out of One, Many are listed under ### Existing Systems; Park et al. 2023 Generative Agents and Gao et al. 2025 AgentSociety are under ### Social Simulation; Park et al. 2024 Generative Agent Simulations of 1,000 People, Aher et al. 2023 Turing Experiments, Horton et al. 2023 Homo Silicus, Törnberg et al. 2023, Toubia et al. 2025 Twin-2K-500, Tencent 2024 1B Personas and Cao et al. 2025 are under ### Synthetic Populations; Meta FAIR 2022 CICERO is under ### Multi-Agent Reinforcement Learning; Chopra et al. 2025 AgentTorch and Zheng et al. 2022 The AI Economist are under ### Agent-Based Modelling; Kıcıman et al. 2023 and Nilforoshan et al. 2024 are under ### Causal Inference. They are deliberately not duplicated here.

Machine-Learned World Models

• Ha & Schmidhuber — World Models (NeurIPS 2018) paper — Google Brain / IDSIA paper introducing learned latent-dynamics world models trained from pixels and using them as the substrate a policy is optimised inside; the foundational reference for "learn the simulator" as a research direction and the canonical antecedent of every later RSSM- and Dreamer-style architecture.
• Hafner, Lillicrap, Fischer, Villegas, Ha, Lee & Davidson — Learning Latent Dynamics for Planning from Pixels (PlaNet, ICML 2019) paper — Google Brain / DeepMind paper introducing the Recurrent State-Space Model (RSSM) that learns a stochastic latent transition model from images and plans inside it; the latent-dynamics primitive almost every modern learned-simulator architecture builds on.
• Hafner, Pasukonis, Ba & Lillicrap — Mastering diverse control tasks through world models (DreamerV3, Nature 2025) paper — Google DeepMind general-purpose world-model agent that masters 150+ tasks across domains with a single set of hyperparameters; the strongest current reference for treating a learned latent-dynamics simulator as a substrate a policy is trained inside rather than alongside.
• Schrittwieser, Antonoglou, Hubert, Simonyan, Sifre, Schmitt, Guez, Lockhart, Hassabis, Graepel, Lillicrap & Silver — Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model (MuZero, Nature 2020) paper — DeepMind model-based RL agent that plans inside a learned environment model without ever being told the rules of the game; the canonical reference for learned-model planning and the proof that simulator-free domains can be controlled through an internal learned simulator.
• Bruce, Dennis, Edwards, Parker-Holder, Shi, Hughes, Lai, Mavalankar, Steigerwald, Apps, Aytar, Bechtle, Behbahani, Chan, Heess, Gonzalez, Osindero, Ozair, Reed, Zhang, Zolna, Clune, de Freitas, Singh & Rocktäschel — Genie: Generative Interactive Environments (ICML 2024) paper — Google DeepMind foundation model trained on internet-scale video that synthesises playable interactive environments from a single image prompt; the strongest current reference for "world model as foundation model" and a demonstration that learned simulators can be conditioned, not just sampled.
• Hu, Russell, Yeo, Murez, Fedoseev, Kendall, Shotton & Corrado — GAIA-1: A Generative World Model for Autonomous Driving (2023) paper — Wayve multi-modal generative world model for driving conditioned on text, image, and action; the canonical reference for steerable generative environments grounded in a real-world social-physical domain rather than a game.
• Valevski, Leviathan, Arar & Fruchter — Diffusion Models Are Real-Time Game Engines (GameNGen, 2024) paper — Google Research diffusion-model neural game engine that simulates DOOM at interactive frame rates conditioned on player input; a pointed demonstration that a neural network alone, with no hand-coded engine, can serve as a steerable real-time simulator.
• Gao et al. — Infinite Worlds with Versatile Interactions (LingBot-World 2.0, 2026) tool — Provides an open-source real-time world-model substrate with long-horizon interactive video generation, multi-user control, and director–pilot agents for coordinating characters and scene evolution.
• Garnelo, Rosenbaum, Maddison, Ramalho, Saxton, Shanahan, Teh, Rezende & Eslami — Conditional Neural Processes (ICML 2018) paper — DeepMind paper introducing the conditional neural process: a neural network mapping a context set of observations to a distribution over functions in a single forward pass; the foundational primitive for treating a neural network as an amortised, callable simulator.
• Garnelo, Schwarz, Rosenbaum, Viola, Rezende, Eslami & Teh — Neural Processes (ICML 2018 workshop) paper — DeepMind latent-variable extension of conditional neural processes that names the "neural process" family and adds calibrated predictive uncertainty; the standard reference cited by every downstream NP variant and the entry point a researcher reaches for to use NPs as uncertainty-aware learned simulators.
• Kim, Mnih, Schwarz, Garnelo, Eslami, Rosenbaum, Vinyals & Teh — Attentive Neural Processes (ICLR 2019) paper — DeepMind extension of neural processes with cross-attention from targets to context that fixes the under-fitting failure mode of vanilla NPs; the practical NP reference for when the primitive needs to actually fit observed data well rather than collapse to a smoothed mean.
• Coletta, Moulin, Vyetrenko & Balch — Learning to Simulate Realistic Limit Order Book Markets From Data as a World Agent (ICAIF 2022) paper — J.P. Morgan AI Research learned market simulator that treats the entire exchange as a single neural "world agent" reactive to order flow; a worked example of a research-lab learned simulator for a real social-economic system and the canonical reference for AI-research-lab market-microstructure simulation.
• Suo, Regalado, Casas & Urtasun — TrafficSim: Learning to Simulate Realistic Multi-Agent Behaviors (CVPR 2021) paper — Uber ATG / University of Toronto neural multi-agent traffic simulator trained on real driving logs that learns reactive, socially plausible behaviour for all road users in a scene; a worked example of a learned simulator applied to a real urban-social system and a methodological template for learned-dynamics simulation of human systems.

Cross-references: Cranmer, Brehmer & Louppe 2020 The Frontier of Simulation-Based Inference, Hermans et al. 2022 A Trust Crisis in Simulation-Based Inference?, Tejero-Cantero et al. 2020 sbi toolkit, Lueckmann et al. 2021 Benchmarking SBI, Dyer et al. 2022 Black-box Bayesian inference for economic ABMs and Talts et al. 2018 Simulation-Based Calibration are listed under ### Uncertainty Quantification; Wang et al. 2023 Voyager is under ### LLM-Based Social Simulation; Meta FAIR 2022 CICERO is under ### Multi-Agent Reinforcement Learning; Zheng et al. 2022 The AI Economist is under ### Agent-Based Modelling. They are deliberately not duplicated here.

Policy and Intervention Modelling

• Capobianco, Kompella, Ault, Sharon, Jong, Fox, Meyers, Wurman & Stone — Agent-based Markov Modeling for Improved COVID-19 Mitigation Policies (JAIR 2021) paper — Sony AI / UT Austin / Texas A&M open-source PandemicSimulator plus a reinforcement-learning policy that searches over fine-grained NPI levers (mask mandates, school closure, contact tracing) to balance ICU load against economic loss; the canonical research-lab worked example of RL-driven non-pharmaceutical-intervention design over a community-scale agent-based COVID simulator.
• Chopra, Rodriguez, Prakash, Raskar & Kingsley — Using neural networks to calibrate agent based models enables improved regional evidence for vaccine strategy and policy (Vaccine 2023) paper — MIT Media Lab / Mayo Clinic HEAL / Georgia Tech / Michigan demonstration that gradient-trained neural networks can calibrate million-agent ABMs in seconds, turning AgentTorch into a regional vaccine-policy decision-support tool; an applied counterpart to the AgentTorch methodology entry under ### Agent-Based Modelling.
• Du, Chen, Yang, Long & Zhao — HRL4EC: Hierarchical Reinforcement Learning for Multi-Mode Epidemic Control (Information Sciences 2023) paper — Jilin University / Nanyang Technological University / Dalian University of Technology hierarchical-RL framework over a Multilateral-Impact-Driven SEIR model that jointly designs contact-restriction, vaccination, and treatment interventions; a non-Western research-lab reference for multi-lever epidemic-policy design that closes the non-Western coverage gap flagged in the project mission.
• Wu, Kreidieh, Parvate, Vinitsky & Bayen — Flow: A Modular Learning Framework for Mixed Autonomy Traffic (IEEE Transactions on Robotics 2022) tool — UC Berkeley deep-RL framework wrapping SUMO that lets researchers train and evaluate control policies for mixed human / autonomous-vehicle traffic; the canonical AI-native substrate for evaluating transport policies (autonomous-vehicle penetration rules, ramp-metering schemes, congestion-pricing experiments) at network scale.
• Wei, Zheng, Yao & Li — IntelliLight: A Reinforcement Learning Approach for Intelligent Traffic Light Control (KDD 2018) paper — Penn State deep-RL controller for traffic signals trained and validated on real-world surveillance-camera traffic data; the most-cited primary reference for RL-driven signal-policy design and the seed of the IntelliLight / FRAP / CityFlow line of city-scale signal-policy work.
• Strnad, Barfuss, Donges & Heitzig — Deep Reinforcement Learning in World-Earth System Models to Discover Sustainable Management Strategies (Chaos 2019) paper — Potsdam Institute for Climate Impact Research demonstration that deep-RL agents acting inside coupled World-Earth system models can discover sustainable-development pathways (mixes of carbon taxation and renewable subsidies) that classical analysis misses; the canonical reference for RL-as-policy-search over learned climate-economy dynamics.
• Zhang, Williams, Phade, Srinivasa, Y. Zhang, Gupta, Bengio & Zheng — AI for Global Climate Cooperation: Modeling Global Climate Negotiations, Agreements, and Long-Term Cooperation in RICE-N (ICML 2025) paper — Mila / Salesforce Research multi-region integrated assessment model with explicit negotiation protocols, used as a multi-agent-RL substrate for designing global climate-cooperation mechanisms; the leading research-lab open environment for AI-for-climate-policy work and the basis of the AI for Global Climate Cooperation competition.
• Dütting, Feng, Narasimhan, Parkes & Ravindranath — Optimal Auctions through Deep Learning: Advances in Differentiable Economics (JACM 2024) paper — Google Research / Harvard EconCS framing of optimal auction design as a constrained deep-learning problem solvable with standard ML pipelines; the canonical reference for differentiable mechanism design and the cross-domain methodology an AI-native policy engine reaches for whenever an intervention can be cast as a mechanism rather than a control loop.

Ethical Risk Discovery

• Pan, Chan, Zou, Li, Basart, Woodside, Ng, Zhang, Emmons & Hendrycks — Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the MACHIAVELLI Benchmark (ICML 2023) paper — UC Berkeley CHAI / Center for AI Safety benchmark of 134 text-game social-decision simulations annotated with power-seeking, deception, and ethical-violation labels; the canonical sociotechnical-eval substrate for measuring harmful behaviour in generative-agent populations before deploying them inside a social-simulation engine.
• Ruan, Dong, Wang, Pitis, Zhou, Ba, Dubois, Maddison & Hashimoto — Identifying the Risks of LM Agents with an LM-Emulated Sandbox (ICLR 2024) tool — Stanford / Toronto ToolEmu framework using a language model to emulate tool execution plus an automatic safety evaluator for LM-agent risks across diverse scenarios; the practical red-teaming methodology a generative-agent simulator's risk layer reaches for to surface unsafe behaviour without manual scenario authoring.
• D'Amour, Srinivasan, Atwood, Baljekar, Sculley & Halpern — Fairness is not Static: Deeper Understanding of Long Term Fairness via Simulation Studies (FAccT 2020) paper — Google paper plus open-source ML-fairness-gym argument that long-term fairness of an algorithmic policy is only legible through simulation of the system it is embedded in; the foundational reference for using simulators specifically to audit the distributional consequences of learned policy over time.
• Liu, Dean, Rolf, Simchowitz & Hardt — Delayed Impact of Fair Machine Learning (ICML 2018) paper — UC Berkeley demonstration that static fairness criteria can produce harm an unconstrained policy would not, visible only when the policy's effect is simulated forward into the population state; the original "fairness-via-simulation" result every learned-policy audit should reproduce as a baseline before claiming distributional safety.
• Heidari, Nanda & Gummadi — On the Long-term Impact of Algorithmic Decision Policies: Effort Unfairness and Feature Segregation through Social Learning (ICML 2019) paper — CMU / MPI-SWS extension of the delayed-impact line that models individuals' effort responses to a deployed policy; the methodological reference for auditing how a learned policy reshapes the population it acts on, including effort-based unfairness an outcome-only audit cannot see.
• Stadler, Oprisanu & Troncoso — Synthetic Data — Anonymisation Groundhog Day (USENIX Security 2022) paper — EPFL / Imperial empirical demonstration that deep-generative synthetic populations do not provide a better privacy-utility trade-off than classical anonymisation and remain vulnerable to membership-inference attacks; the load-bearing reference for the synthetic-population leakage risk an AI-native social-simulation engine inherits the moment it ships its synthetic society.
• Park, Goldstein, O'Gara, Chen & Hendrycks — AI Deception: A Survey of Examples, Risks, and Potential Solutions (Patterns 2024) paper — MIT-led survey cataloguing learned-deception behaviour by AI systems trained inside game and negotiation simulators (CICERO in Diplomacy, social-deduction agents, multi-agent bargaining); the canonical reference for the dual-use / second-order-harm pathway where generative agents inside a social-simulation engine learn deception that transfers to deployment.
• Solaiman, Talat, Agnew, Ahmad, Baker, Blodgett, Daumé III, Dodge, Hooker, Jernite, Mitchell, Newman, Strait et al. — Evaluating the Social Impact of Generative AI Systems in Systems and Society (2023) framework — Hugging Face / Stanford / Allen AI / multi-lab categorical framework dividing generative-AI evaluation into base technical-system properties (bias, performance disparities, privacy, environmental and financial costs) and societal-context evaluations (trustworthiness, distributional harms, labour-market effects); the off-the-shelf taxonomy a learned simulation engine's risk-discovery layer reuses to enumerate what it must measure on its generative-agent and learned-policy outputs.

Datasets and Empirical Grounding

• Panel Study of Income Dynamics (PSID, 1968–) dataset — University of Michigan Institute for Social Research longest-running household panel in the world, biennial since 1968 covering income, wealth, intergenerational transfers, employment, and health for ~9,000 US families with 8,000+ children of original participants now followed; the load-bearing reference for multi-decade validation of intergenerational dynamics in microsimulation, learned-world-model engines, and the Stanford / MIT / NBER computational-economics simulator stack. Access: open with online registration via the PSID Data Center.
• National Longitudinal Study of Adolescent to Adult Health (Add Health, 1994–) dataset — UNC Carolina Population Center longitudinal panel of ~20,000 US adolescents followed into adulthood across five waves with health, behaviour, peer-network, biomarker, and genetic modules; the canonical substrate Harvard / MIT / Stanford computational-social-science network simulators (Christakis–Fowler peer-effects line, SNAP-style network-dynamics work) reach for to ground adolescent peer-influence and social-contagion dynamics in agent-based and learned-network models. Access: open public-use file via ICPSR; restricted-use sensitive data on-application.
• Health and Retirement Study (HRS, 1992–) dataset — University of Michigan Institute for Social Research biennial longitudinal panel of ~20,000 US adults aged 50+ with deep modules on health, cognition, biomarkers, finance, and family transfers; the canonical aging / health / mortality calibration reference for US-population simulators and the anchor of the HRS-international sister-study triad NBER, Stanford SIEPR, and aging-policy ML labs cite as the gold-standard substrate for retirement, long-term-care, and health-shock dynamics. Access: open with registration; sensitive linkage data on-application.
• China Health and Retirement Longitudinal Study (CHARLS, 2011–) dataset — Peking University National School of Development biennial longitudinal panel of ~17,000 Chinese adults aged 45+ across multiple waves with health, biomarkers, cognition, family, and finance modules; a non-Western canonical reference for aging, health, and mortality calibration and the China arm of the HRS-international sister-study triad cross-cultural aging-policy ML work cites for non-Western validation. Access: on-application via the CHARLS data portal.
• IPUMS USA — Integrated Public Use Microdata Series, USA (1850–) dataset — Minnesota Population Center research-active redistribution and harmonisation of US Census Bureau decennial census and American Community Survey microdata back to 1850, all variables harmonised to consistent codes; the canonical empirical input for US synthetic-population grounding and the data layer underneath the Lovelace 2024 country-scale synthetic population already curated above. Access: open with registration.
• IPUMS International — Integrated Public Use Microdata Series, International (1960–) dataset — Minnesota Population Center research-grade harmonisation of census microdata across 104 countries with extensive non-Western coverage and consistent variable codes across decades; the canonical multi-country empirical input for cross-country synthetic-population grounding outside the United States and the substrate cross-cultural population-engine and migration-simulation work cites for non-Western synthesis. Access: open with registration and approved-use agreement.
• CHARLS — China Health and Retirement Longitudinal Study (2011–) dataset — Peking University biennial longitudinal panel of ~17,000 Chinese adults; the canonical non-Western reference for aging, health, and mortality calibration in social engines.
• Kashiyama et al. — Pseudo-PFLOW: Nationwide Synthetic Human Mobility Dataset Construction from Limited Travel Surveys and Open Data (2024) paper — Tokyo/Osaka research-lab synthetic population covering all ~130 million people in Japan; closes the non-Western coverage gap for country-scale AI-native population builds.
• General Social Survey (GSS, 1972–) dataset — NORC at the University of Chicago canonical US attitudinal and social-behaviour survey across 30+ waves; the validation substrate Park et al. 2024 Generative Agent Simulations of 1,000 People (Stanford + Google DeepMind) used to test whether LLM-driven agents replicate participants' GSS responses at 85% of test–retest reliability — the de facto AI-native benchmark for generative-agent fidelity to a real US population sample. Access: open download via the GSS Data Explorer.
• American National Election Studies (ANES, 1948–) dataset — University of Michigan + Stanford long-running US political-attitudinal survey covering presidential and midterm cycles with deep ideological, demographic, and behavioural modules; the silicon-sample substrate Argyle et al. 2023 Out of One, Many: Using Language Models to Simulate Human Samples (Stanford / BYU) used to demonstrate LLM-conditioned political-attitudinal distributions can match real ANES respondent distributions, and the political-opinion benchmark DeepMind / Stanford LLM-society simulators reach for to calibrate ideological dynamics. Access: open download with registration.
• OpinionQA — Whose Opinions Do Language Models Reflect? (Santurkar et al. ICML 2023) dataset — Stanford NLP / Tatsu Hashimoto group AI-native opinion-benchmark dataset built from Pew Research American Trends Panel (~1,500 questions × 60+ topics × demographic subgroups), released alongside the ICML 2023 paper; the canonical AI-native substrate the Whose Opinions line uses to measure LLM ideological bias and underrepresentation of demographic subgroups, reused across Anthropic / DeepMind / Stanford LLM-alignment evaluation. Access: open download via the project GitHub repository (with redistributed Pew ATP excerpts).
• PRISM Alignment Project (Kirk et al. NeurIPS Datasets & Benchmarks 2024) dataset — Oxford Internet Institute + Cohere + MetaAI + Hugging Face + multi-lab participatory-alignment dataset of 8,011 live conversations between 1,500 participants from 75 countries with 21+ values dimensions and stated preferences over LLM responses; the canonical multi-stakeholder substrate value-pluralism and cross-cultural LLM-alignment work (constitutional-AI variants, multilingual feedback, generative-agent value-diversity) cites for grounding alignment beyond US-WEIRD samples. Access: open download via Hugging Face Datasets under the project licence.
• World Values Survey (WVS, 1981–) dataset — World Values Survey Association (Vienna) cross-cultural attitudinal survey covering 100+ countries across seven completed waves with strong non-Western coverage (East Asia, MENA, sub-Saharan Africa, Latin America); the canonical reference Anthropic GlobalOpinionQA, DeepMind cross-cultural LLM-society work, and Stanford generative-agent calibration papers reach for to validate non-Western value-distribution priors and test cross-cultural transfer beyond English-speaking populations. Access: open download with registration.
• MovieLens (1995–) dataset — GroupLens Research at the University of Minnesota canonical movie-rating datasets ranging from 100k ratings (small benchmark) to 33M ratings (full corpus) collected from active MovieLens users; the load-bearing substrate Microsoft Research / Meta / Google recommender-simulation work, learned-user-model evaluation, and counterfactual-bandit research reach for as the de-facto benchmark for behavioural-trace simulation of preference dynamics. Access: open download under research licence.

Tools and Libraries

• JAX (Google DeepMind, 2018–) tool — Google DeepMind functional-array library combining NumPy-compatible APIs with autodiff (grad), JIT compilation (jit), automatic vectorisation (vmap), and parallelisation (pmap) over XLA; the AI-lab-canonical accelerator substrate underneath DeepMind's Concordia generative-agent engine (curated in ### Existing Systems), Brax differentiable physics, the Melting Pot 2.0 multi-agent suite (curated in ### Multi-Agent Reinforcement Learning), and the NumPyro / BlackJAX probabilistic-programming stack Stanford and academic Bayesian-calibration groups use for AI-native simulators. Python; production-grade actively maintained.
• Paszke et al. — PyTorch: An Imperative Style, High-Performance Deep Learning Library (NeurIPS 2019) tool — Meta FAIR deep-learning framework with imperative tape-based autograd, dynamic graph construction, and a CUDA backend; the load-bearing deep-learning substrate Anthropic Claude training, Stanford NLP generative-agent work, Hugging Face Transformers, Meta FAIR CICERO Diplomacy, Stable-Baselines3, and the broader academic AI-native simulation literature run on. Python with C++ kernels; production-grade.
• Wolf et al. — Transformers: State-of-the-Art Natural Language Processing (EMNLP 2020 demo) tool — Hugging Face open-source library hosting tens of thousands of pre-trained transformer checkpoints with unified loading, fine-tuning, and inference APIs; the LLM scaffolding underneath Park et al. Generative Agent Simulations of 1,000 People (Stanford + Google DeepMind), the Yang et al. OASIS million-agent simulator (curated in ### Existing Systems), and the silicon-sample / synthetic-population LLM workflows top labs reach for. Python (PyTorch / JAX / TensorFlow backends); production-grade.
• Abril-Pla et al. — PyMC: A Modern, and Comprehensive Probabilistic Programming Framework in Python (PeerJ Computer Science 2023) tool — PyMC Labs / NumFOCUS-governed open-source probabilistic-programming library with declarative model specification, NUTS / SMC / variational-inference samplers, and a JAX backend via PyTensor; the calibration substrate Oxford INET economic-ABM groups, Berkeley computational-social-science labs, and academic generative-Bayesian-microsimulation research reach for to fit posterior distributions over ABM and microsimulation parameters when the simulator likelihood is tractable. Python; production-grade actively maintained.
• Phan, Pradhan & Jankowiak — Composable Effects for Flexible and Accelerated Probabilistic Programming in NumPyro (2019) tool — Uber AI Labs (now open-source) JAX-backed probabilistic-programming library that mirrors the Pyro effect-handler API on top of JAX's jit / grad / vmap stack; the AI-lab-canonical Bayesian-inference substrate when JAX is already in the deep-learning core, used by Stanford, DeepMind-adjacent academic groups, and the broader Bayesian-deep-learning community for accelerated MCMC and SVI on large social simulators. Python (JAX); research-grade actively maintained.
• Iwanaga, Usher & Herman — Toward SALib 2.0: Advancing the accessibility and interpretability of global sensitivity analyses (Socio-Environmental Systems Modelling 2022) tool — Open-source academic Python library implementing Sobol, Morris, FAST, delta, and PAWN global sensitivity-analysis methods behind a unified sampler-and-analyser API; the canonical sensitivity-analysis substrate Oxford INET economic-ABM groups, academic agent-based-modelling literature, and learned-simulator calibration pipelines use to surface which simulator parameters drive which outcomes before committing to expensive Bayesian calibration. Python; research-grade actively maintained.
• Kumar, Carroll, Hartikainen & Martin — ArviZ a unified library for exploratory analysis of Bayesian models in Python (JOSS 2019) tool — NumFOCUS-governed open-source library for Bayesian-model exploratory analysis (posterior plots, MCMC convergence diagnostics R-hat / ESS, leave-one-out cross-validation, model comparison via PSIS-LOO and WAIC); the cross-platform diagnostic layer the Gelman / Vehtari Stanford–Aalto Bayesian-Workflow line (the canonical methodology paper for Bayesian-simulator validation) prescribes and the Oxford INET economic-ABM Bayesian-calibration pipelines, sbi-toolkit users (curated in ### Uncertainty Quantification), and Stanford / Berkeley generative-Bayesian work depend on for posterior diagnostics before declaring a simulator calibrated. Python; production-grade actively maintained.
• Raffin, Hill, Gleave, Kanervisto, Ernestus & Dormann — Stable-Baselines3: Reliable Reinforcement Learning Implementations (JMLR 2021) tool — German Aerospace Center (DLR) / community open-source PyTorch implementations of the canonical model-free RL algorithms (PPO, SAC, TD3, A2C, DQN) with rigorously tested baselines and PettingZoo / Gymnasium compatibility; the de-facto PPO / SAC baseline cited by the Oxford FLAIR / WhiRL MARL line (QMIX baseline comparisons curated in ### Multi-Agent Reinforcement Learning), the Hugging Face Deep RL Course pedagogy (millions of citations through the huggingface_sb3 integration), and the broader academic cooperative-MARL literature when reporting how a proposed method compares to standard RL on social-dilemma and cooperative-coordination environments (complement to RLlib curated in ### Multi-Agent Reinforcement Learning). Python (PyTorch); production-grade actively maintained.
• Hoffman et al. — Acme: A Research Framework for Distributed Reinforcement Learning (2020) tool — Google DeepMind modular RL agent library with JAX (acme.jax) and TensorFlow backends, distributed actor / learner abstractions, and reference implementations of D4PG, IMPALA, MPO, R2D2, and other DeepMind-line algorithms; the in-house DeepMind RL substrate and the canonical reference for distributed RL research at the AI-lab scale, sitting alongside the Melting Pot 2.0 evaluation suite curated in ### Multi-Agent Reinforcement Learning. Python (JAX / TF); research-grade actively maintained.
• Hagberg, Schult & Swart — Exploring Network Structure, Dynamics, and Function using NetworkX (SciPy 2008) tool — Originally Los Alamos National Lab, now NumFOCUS-governed open-source Python network-science library with extensive graph-theoretic algorithms, generative-network models, and centrality / community-detection routines; the load-bearing network-primitive substrate Stanford SNAP-style research, Harvard / MIT computational-social-science network simulators, and Add Health-grounded peer-influence work (curated in ### Datasets and Empirical Grounding) use to construct, mutate, and analyse social networks underneath their simulators. Python; production-grade actively maintained.
• Jordahl et al. — GeoPandas: Python tools for geographic data tool — Open-source academic Python library extending pandas DataFrames with geospatial-vector geometry columns, projection-aware spatial joins, and spatial indexing on top of Shapely / pyproj / GDAL; the geospatial-primitive substrate underneath the Lovelace 2024 country-scale geographically-explicit US synthetic population (curated in ### Microsimulation), the Bilal et al. CitySEIRCast city-scale digital twin (curated in ### Synthetic Populations), and the broader geographically-grounded ABM literature. Python; production-grade actively maintained.
• Moritz et al. — Ray: A Distributed Framework for Emerging AI Applications (OSDI 2018) tool — UC Berkeley RISELab / Anyscale Apache-2 distributed-computing framework with first-class Python actors, task scheduling, and an ecosystem of higher-level libraries (RLlib already curated in ### Multi-Agent Reinforcement Learning, Tune for hyperparameter search, Train for distributed training); the canonical substrate AI labs and academic groups use to distribute simulation runs, RL training, and Bayesian-calibration sweeps across compute clusters. Python with C++ kernels; production-grade actively maintained.
• Lhoest et al. — Datasets: A Community Library for Natural Language Processing (EMNLP 2021 demo) tool — Hugging Face open-source library providing one-line load_dataset(...) access to tens of thousands of ML datasets with memory-mapped backing, streaming, and reproducible versioning; the canonical loader infrastructure underneath the OpinionQA, PRISM Alignment Project, and behavioural-trace substrates curated in ### Datasets and Empirical Grounding, used pervasively by Anthropic, Stanford NLP, Cohere, and Hugging Face research code. Python; production-grade actively maintained.
• Biewald — Experiment Tracking with Weights and Biases (2020) tool — Weights & Biases Inc. research-grade experiment-tracking and ML-ops platform (open-source client + free-academic-tier hosted backend) with automatic metric logging, hyperparameter sweeps, artefact lineage, and run comparison; the experiment-management substrate cited pervasively in Anthropic safety / interpretability work, Stanford NLP generative-agent papers, MIT / Hugging Face research code, and academic AI-native simulator reproducibility appendices. Python; production-grade.
• Yadan — Hydra: A framework for elegantly configuring complex applications (Meta FAIR, 2019–) tool — Meta FAIR-originated open-source Python configuration framework with hierarchical YAML composition, command-line overrides, and config validation via OmegaConf / structured configs; the canonical config-management layer underneath PyTorch Lightning research code, Meta FAIR open-source releases (CICERO, Llama-related research scaffolding), and academic AI-lab simulation pipelines that need reproducible per-run configuration sweeps. Python; production-grade actively maintained.

Benchmarks and Testbeds

• Agapiou et al. — Melting Pot 2.0 (DeepMind 2023) tool — DeepMind's flagship evaluation suite for multi-agent reinforcement learning; provides 50+ complex social scenarios to test generalisation of cooperation, trust, and reciprocation in population-scale simulations.
• Google DeepMind — Kaggle Game Arena (2025) framework — A persistent, dynamic benchmarking platform using games like Werewolf and Poker to evaluate AI agents' ability to reason under uncertainty, detect deception, and manage hidden information.
• Froger et al. — Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments (Meta SuperIntelligence Labs 2026) framework — A realistic, asynchronous benchmark where environments evolve independently of agent actions; tests temporal reasoning, ambiguity handling, and noise tolerance in complex agentic workflows.
• Anthropic — Agentic Misalignment Evaluations (2025) paper — A high-fidelity red-teaming benchmark using simulated corporate environments to surface risky autonomous behaviours — such as blackmail or whistleblowing — when AI goals conflict with organizational constraints.
• Chuang et al. — DEBATE: A Large-Scale Benchmark for Role-Playing LLM Agents in Multi-Agent, Long-Form Debates (NeurIPS 2025) dataset — The first large-scale empirical benchmark for evaluating the authenticity of multi-agent group dynamics; compares LLM-simulated opinion trajectories against data from 2,792 human participants.
• Isomorphic Labs — Isomorphic Labs Drug Design Engine (IsoDDE) Technical Report (2026) framework — A unified computational system surpassing AlphaFold 3 in protein-ligand structure prediction and binding affinity; introduces the 'Runs N' Poses' benchmark for testing generalisation to novel biological systems.
• Google Research — CURIE: Evaluating LLMs on Multitask Scientific Long-Context Understanding and Reasoning (ICLR 2025) paper — A multitask benchmark across six scientific disciplines (including materials science and geospatial analysis) that tests long-context reasoning, information extraction, and aggregation in realistic scientific workflows.
• Pan et al. — MACHIAVELLI Benchmark (ICML 2023) paper — A benchmark of 134 text-game social-decision simulations designed to measure trade-offs between utility-seeking and ethical behaviour, specifically surfacing power-seeking and deceptive tendencies in agents.

Bounded Slices and Composability

• Pogonyi — Transport Modeling (2020) chapter — Surveys how bounded transport models serve as tool-models for urban economic planning.
• Mérő, Borsos, Hosszú, Oláh & Vágó — A high-resolution, data-driven agent-based model of the housing market (2023) paper — A self-contained slice modelling the interaction between credit constraints and housing prices.
• Flache et al. — Models of Social Influence: Towards the Next Generation (2017) paper — A foundational review of bounded models for opinion dynamics and norm formation.
• Lorenz — Continuous Opinion Dynamics under Bounded Confidence: A Survey (2007) paper — The definitive reference for the "bounded confidence" slice of social influence modelling.
• Althouse et al. — Enhancing disease surveillance with novel data streams: challenges and opportunities (2015) paper — Explores how web-search and social-media-derived behaviour slices inform epidemic surveillance and simulation engines.
• Adam — Special report: The simulations driving the world’s response to COVID-19 (2020) article — Nature guide to the "mechanistic slice" approach in epidemiology.
• Acemoglu & Jackson — History, Expectations, and Leadership in the Evolution of Social Norms (2011) paper — MIT work modelling how individual beliefs and expectations stabilize or shift cooperative norms; provides a template for modelling institutional behaviour beyond individual agents.
• Golder & Macy — Digital Footprints: Opportunities and Challenges for Online Social Research (2014) paper — Explores how fine-grained, time-stamped digital traces can support empirical validation of social dynamics in the digital age.
• Cao — Research on co-simulation of multi-resolution models based on HLA (2023) paper — Technical patterns for orchestrating multiple simulation engines through federated model coupling and shared simulation state.
• Gao et al. — Large language models empowered agent-based modeling and simulation: a survey and perspectives (2024) paper — Surveys the "simulation-as-a-tool" pattern for LLM-driven agent-based simulation workflows.

Frontier Science Parallels

• Jumper et al. — Highly accurate protein structure prediction with AlphaFold (2021) paper — The landmark reference for geometric deep learning and structural priors in complex system simulation; establishes the precedent for "learned simulation" of large-scale interactions.
• Isomorphic Labs — The Isomorphic Labs Drug Design Engine (2026) article — Introduces the unified computational system for protein-ligand structure prediction; a methodological cousin to the social-system "tool model" ecosystem.
• Pathak et al. — FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators (2022) paper — Demonstrates 1,000x speedups in weather simulation via learned latent dynamics; a direct parallel for social world models grounded in massive trace data.
• Kochkov et al. — Machine learning–accelerated computational fluid dynamics (2021) paper — Shows how neural networks can accelerate mechanistic physical simulations; provides a template for neural surrogates in social agent-based modelling.
• Merchant et al. — Scaling deep learning for materials discovery (GNoME, 2023) paper — Google DeepMind work using active learning to discover 2.2 million new crystals; the canonical reference for large-scale, automated search over simulation-generated probability landscapes.
• Karr et al. — A whole-cell computational model predicts phenotype from genotype (2012) paper — Canonical virtual-cell work showing how mechanistic submodels can be composed into a larger simulator; a concrete methodological parallel for social-system modelling.
• Chen, Werling & Liu — Real-time Model Predictive Control and System Identification Using Differentiable Physics Simulation (2022) paper — A reference for differentiable simulators that allow gradient-based policy optimization; the technical bridge to differentiable social engines.

Transferability & Constraints: While scientific simulation provides patterns for latent dynamics and scalability, social systems differ in their observability limits (hidden human intent), ethical risks (distributional harm), and validation boundaries (the lack of invariant laws like gravity or thermodynamics).

Implementation Patterns

• Chopra et al. — flame: A Framework for Learning in Agent-based ModEls (AAMAS 2024) tool — The AgentTorch implementation of differentiable ABMs in PyTorch; enables automatic differentiation through agent interactions for rapid calibration.
• Andelfinger — Towards Differentiable Agent-Based Simulation (2022) paper — Demonstrates how automatic differentiation can support gradient-based optimization and calibration in complex social ABMs.
• Dyer et al. — Interventionally Consistent Surrogates for Complex Simulation Models (2024) paper — Oxford INET framework for learning neural surrogates that remain consistent under counterfactual perturbations.
• Hafner et al. — Mastering diverse control tasks through world models (DreamerV3, Nature 2025) paper — The reference architecture for latent dynamics world models (RSSM) that can serve as the learned substrate for social simulation.
• Tejero-Cantero et al. — sbi: A toolkit for simulation-based inference (2020) tool — The industry-standard Python library for amortised neural posteriors (NPE/NLE/NRE) on simulation outputs.
• Cranmer et al. — The Frontier of Simulation-Based Inference (2020) paper — The foundational methodological framework for neural SBI, providing the blueprint for uncertainty-aware learned simulators.

Evaluation and Validation

• Talts et al. — Validating Bayesian Inference Algorithms with Simulation-Based Calibration (2018) paper — The standard diagnostic for verifying whether a calibration pipeline is itself statistically well-calibrated.
• Salter et al. — Improving Policy-Oriented Agent-Based Modeling with History Matching (2025) paper — Demonstrates history matching as a principled method for calibrating simulation engines against real-world data traces.
• Mei et al. — A Turing test of whether AI chatbots are behaviorally similar to humans (2024) paper — A landmark empirical anchor for evaluating the behavioural fidelity of synthetic agents in social tasks.
• Grimm et al. — The ODD Protocol for Describing Agent-Based and Other Simulation Models (2020) framework — The international standard for model documentation, ensuring technical replicability and structural transparency.
• Lueckmann et al. — Benchmarking Simulation-Based Inference (2021) paper — A comprehensive benchmark suite for choosing the right calibration algorithm for high-dimensional social simulators.
• Park et al. — Generative Agent Simulations of 1,000 People (2024) paper — Establishes the 85% test-retest reliability benchmark for synthetic agents against real US population survey data.

Standards and Reproducibility

• VISA authors — VISA: A Structured Description Protocol for Agent-Based Simulation Models Towards Machine Reproducibility (2026) framework — Defines a structured protocol for machine-readable agent-based model descriptions that supports automated model reconstruction and reproducible simulation.

Contributing

Thrilled to have you here.
Whether it's a quick typo fix, a fresh resource,
a doc polish, or a sweeping overhaul — every contribution helps this list grow.
Jump in and join the community — PRs of every size are welcome.

📝 Read the contributing guide · 🐛 good first issues