Grounded Social Environment Design

Modern public systems run on a feedback loop. Institutional policies perturb how people behave; millions of decentralized decisions aggregate and reshape the physical and social state of the world; and the changed world forces planners to revise their policies. That loop is exactly what makes policy hard to evaluate in advance: a randomized controlled trial on a real population is slow, expensive, and often ethically fraught, and by the time it reports, the population it measured no longer exists. Large language models, trained on vast human corpora, have emerged as a promising substrate for simulating human decision-making without hand-coded heuristics^1,21  Park et al., 2023. Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023.2  OASIS Authors, 2024. Open Agent Social Interaction Simulations with One Million Agents. arXiv 2024. — but today’s simulators are ad-hoc, domain-specific systems that treat physical interactions and cognitive attitudes in isolation, collapsing individual demographic variance into homogenized, population-average profiles.

This post describes Grounded Social Environment Design (G-SED), our proposed answer: a unified, domain-agnostic, four-layer architecture in which high-fidelity, attribute-conditioned LLM agents live inside a closed-loop Stackelberg Markov game, co-adapting across physical action boundaries and attitudinal opinion spaces at once. It is a research proposal in the strict sense — an architecture, a training pipeline, and a validation plan with the measurement bar committed to before any results — and this post lays out all three.

The two halves of policy simulation

Automated policymaking has progressed along two mature tracks that have never been connected. The first is mechanistic environmental design: frameworks such as the LLM Economist,³3  Karten et al., 2025. LLM Economist: Large Population Models and Mechanism Design in Multi-Agent Generative Simulacra. arXiv 2025. Social World Models,⁴4  Zhang, Huang et al., 2025. Social World Model-Augmented Mechanism Design Policy Learning. NeurIPS 2025. and adaptive agent-based modelling (ADAGE)⁵5  Evans et al., 2025. ADAGE: A Generic Two-Layer Framework for Adaptive Agent-Based Modelling. AAMAS 2025. use multi-agent reinforcement learning or deep world models to optimize structural rules: tax policies, spatial layouts, market parameters. The loop is closed — agents genuinely adapt — but the agents themselves are mathematically rigid utility functions. No cognitive structure, no memory decay, no social identity. They cannot simulate policy acceptability, compliance gaps, or psychological resistance, which is where real policies actually fail.

The second track is static distribution-aligned survey modeling: platforms like Stanford’s Socrates⁶6  Kolluri et al., 2025. Finetuning LLMs for Human Behavior Prediction in Social Science Experiments. arXiv 2025. and subpopulation survey predictors⁷7  Suh et al., 2025. Language Model Fine-Tuning on Scaled Survey Data for Predicting Distributions of Public Opinions. ACL 2025. fine-tune language models to replicate aggregate human opinion distributions, and audits ask precisely whose opinions those models reflect.⁸8  Santurkar et al., 2023. Whose Opinions Do Language Models Reflect? ICML 2023. These systems capture realistic demographic and ideological variance — and then answer exactly once. They are one-shot instruments with no state transitions, no way to express how attitudes evolve when a policy changes and the environment feeds back.

We formalize both tracks as ends of a single spectrum, extending the position that social environment design should be developed for AI-based policymaking.⁹9  Zhang et al., 2024. Position: Social Environment Design Should be Further Developed for AI-based Policy-Making. ICML 2024. At one end lie physical-action environments — spatial coordination, resource extraction, traffic bottlenecks — governed by coordinate geometry and resource constraints. At the other lie socio-cognitive environments — media polarization, institutional trust, regulatory acceptability — governed by symbolic persuasion, subjective beliefs, and cultural narratives.

The cost of leaving that middle empty has a name: the Lucas Critique. When a planner changes physical parameters — tax structures, lockdowns, enforcement cameras — the change alters not just the agents’ immediate actions but the underlying cognitive rules, beliefs, and behavioral elasticities of the population. A simulator whose people cannot change their minds will validate policies that fail the moment real people do.

The closed loop, formally

G-SED formalizes policy synthesis as a bi-level optimization problem: a Stackelberg Markov game S = (n, Φ, Π, u) between a central planner (the leader) and a population of n heterogeneous agents (the followers). The leader commits to a structured policy parameter φ, which parameterizes a lower-level n-player partially observable Markov game. The followers observe the policy signal through individual cognitive-belief filters — not directly, which matters — and respond with an adaptive best-response profile π*(φ). The leader’s objective is the aggregate social welfare of the whole arrangement, subject to budget:

max_{φ ∈ Φ}  u₀(φ, π*(φ))   s.t.   Σ_d cost(d) · intensity(d) ≤ B

while each follower maximizes its own expected long-term return inside the game the policy defines. The interesting move is that the follower response is operationalized in two parallel modes that share one aggregation and policy engine. In Game Mode, agents act in sequential multi-agent coordination games — movement, resource harvesting, strategic bidding — and the environment analytically computes an expected damage or payoff score per agent from physical constraints and spatial interactions. In Survey Mode, the same agents are presented with structured questionnaires and produce ordinal Likert-scale responses, mapping their cognitive and attitudinal profiles into expressed public opinion. One population, one policy signal, two read-outs: what people do and what they say.

Beliefs that move

The piece that makes the loop mean something is dynamic belief propagation. When the planner deploys a revised policy, agents do not respond from a blank slate. Each agent carries internal belief states — an enforcement belief (how likely violations are to be caught) and a risk perception (how dangerous non-compliance feels) — that update through nonlinear transition functions when the policy changes, and then decay back toward the agent’s historical prior:

EB_i(t+1) = (1 − λ_i) · clip( EB_i(t) · φ′/φ + α · Δφ · (1 − d_i), 0, 1 ) + λ_i · EB_i(0)

Three individual parameters do the psychological work: λ_i is the agent’s memory decay rate — how quickly the new enforcement reality fades back toward old habits; d_i is an intrinsic detection bias; and r_i, in the twin risk-perception update, is a risk-denial score. A population defined this way does not just have a distribution of demographics; it has a distribution of forgetting, which is exactly the property a compliance policy lives or dies on. This is the Lucas Critique turned into state variables: the behavioral elasticities are no longer constants the simulator assumes, they are quantities the simulation tracks.

The cognitive engine

None of this works if the agents themselves are thin wrappers around a commercial API. Off-the-shelf assistants suffer cross-persona contamination, carry measured demographic bias,⁸ and cap the token budgets a million-inference simulation needs. So G-SED’s inference layer is the Attribute-Conditioned Synthetic Agent Foundation (ACSA-F): an open, high-capacity model — Qwen-2.5-14B-Instruct in the reference implementation — put through a three-stage behavioral fine-tuning pipeline that forces its representational layers to segregate along socio-demographic, psychographic, and behavioral boundaries.

Stage one: supervised fine-tuning for joint distribution alignment. The base model is tuned on a merged corpus of individual-level social science data, conditioning each response on explicit demographic and psychographic attributes. Two datasets anchor it: SocSci210 — 2.9 million individual human responses from 400,491 participants across 210 peer-reviewed NSF TESS experiments⁶ — and SubPOP, 70,000 subpopulation-response pairs across 3,362 public opinion questions from Pew ATP and NORC GSS, with individual responses sampled from the empirical subpopulation distributions to calibrate the model’s token likelihoods.⁷

Stage two: oracle chain-of-thought. To stop the model from issuing immediate, superficial choices, step-by-step cognitive reasoning traces are prepended before the choice token — drawn from datasets that already carry human qualitative justifications, and from teacher-model distillation under strict prompts requiring three to five sentences explaining why a specific subpopulation identity holds this attitude. A filtering gate drops any generated trace that fails an automated keyword-grounding check against the active persona.

Stage three: preference optimization with contrastive persona swaps. Standard fine-tuning drifts distinct personas toward a centrist, helpful “average assistant” profile.⁸ The counter-move: for an identical question, build preference pairs where the preferred answer is the target persona’s empirical choice and the dispreferred answer is the choice of an opposing persona — one far away in age, location, and social value orientation. Direct Preference Optimization on these pairs explicitly rewards inter-group differentiation. The model is punished, in the loss, for making everyone sound the same.

At inference time the engine is strictly bounded: every response must be a deterministic, JSON-wrapped record matching a typed behavior schema — a reasoning array, an ordinal stance index, and an internal confidence scale. The simulator’s per-agent loop is short enough to state in full:

FOR EACH agent i of N:
    read belief states           EB_i(t), RP_i(t)
    build prompt                 persona_i + policy φ_t + question + beliefs
    sample                       y ← vLLM(prompt, temperature 0.2)
    IF schema check fails:       y ← deterministic prior for persona_i
    record                       {stance, compliance, risk} → aggregation

The fallback line matters: a malformed generation degrades to the persona’s baseline prior instead of crashing the run or silently injecting noise — structural parsing safety as a first-class design constraint.

Four layers, typed boundaries

The architecture that wraps this engine is deliberately boring in the same way good infrastructure is. G-SED decouples social simulation into four layers with standardized, schema-validated interfaces between them:

The termination rule is explicit. Each iteration computes a domain-specific system harm — a function of quantities like a congestion or risk index, immunity gaps, or polarization — and the loop ends when the harm metric changes by less than a tolerance for k consecutive iterations, when the budget exhausts, or when the iteration cap is reached. Related closed-loop efforts — LLM-powered social digital twins for population-level policy response¹¹11  Social Digital Twins Authors, 2026. LLM-Powered Social Digital Twins: A Framework for Simulating Population Behavioral Response to Policy Interventions. arXiv 2026. — share the ambition; G-SED’s contribution is the typed, domain-agnostic decomposition plus the attribute-conditioned engine underneath it.

The bar it has to clear

A proposal that simulates society owes its readers the standard it will be judged by, in advance — researchers have argued forcefully that strong preconditions must be met before simulations touch policy.¹³13  Wang et al., 2026. We Need Strong Preconditions for Using Simulations in Policy. PoliSim Workshop, CHI 2026. G-SED adopts the RADIUS evaluation suite on a held-out test split, with three gates every fine-tuned agent population must pass before any policy conclusion is drawn:

The third gate is the honest one. The Wasserstein bound is not measured against zero — it is measured against the empirical best: the human-to-human bootstrap variance, the disagreement real subpopulations have with themselves. A simulated population is not asked to be more consistent than humanity; it is asked to sit within 10% of humanity’s own noise floor.

A simulator whose people cannot change their minds will validate policies that fail the moment real people do. The bar is not realism in the abstract — it is matching human subpopulations, gate by gate, before a single policy conclusion ships.

Where it runs

The validation paradigm is deliberately multi-domain, because a domain-agnostic claim dies on its first domain. Three are scoped: transport (spatial coordination and congestion under pricing and enforcement policies, squarely in Game Mode), public health (vaccination and compliance, where disease physics meets institutional trust — both modes at once), and platform governance (moderation and information policy, squarely in Survey Mode’s attitudinal space). The implementation plan is open-source end-to-end, which is the only credible posture for a system whose purpose is to be argued with.

Coda

Every policy simulator embeds a theory of people. The reinforcement-learning tradition embeds a flattering one: people as tireless optimizers of fixed utility. The survey tradition embeds a frozen one: people as a snapshot of last year’s opinions. Both fail the same way — the moment a policy changes who people are, the simulator is describing a population that no longer exists. G-SED’s wager is that the fix is architectural: put a demographically faithful cognitive engine inside a closed loop, let beliefs update and decay like beliefs do, type every boundary, and refuse to report a result until the agents match the humans they claim to be. Policies perturb people; people reshape policies. A testbed that ignores either half isn’t a testbed — it’s a rehearsal for being wrong at scale.

@misc{ghosh2026gsed,
  author = {Ghosh, Basab and Katte, Abhijeet},
  title  = {Grounded Social Environment Design: A Generalized Framework
            for Attribute-Conditioned LLM Agent Simulation in
            Closed-Loop Synthetic Societies},
  year   = {2026},
  month  = {July},
  url    = {https://www.cmul8.com/research/grounded-social-environment-design/}
}