S²-MAS: 12M Household Simulation at 37K Tokens/s, 421M YN-mini Beats Jev Model

Overview

Hong Kong University's S²-MAS research team, led by Professor Ye Luo, presents a society-scale multi-agent simulation framework that pushes household simulation to 12 million heterogeneous agents using only 14 GPUs (8×H100 80GB + 6×A100 40GB) in under 4 hours. The framework comprises three core components: a 500-million household persona library, the lightweight YN-mini decision model family, and the high-performance Genese simulation engine. The work demonstrates that specialized small models can surpass large general-purpose models in domain-specific simulation accuracy while drastically reducing compute costs.

Data: 500M Household Personas with 36 Structured Fields

Each persona represents a household financial decision-maker with 36 structured fields covering demographics, household structure, income, expenditure, assets, debts, and financial attitudes. Fields are numerically encoded for efficient filtering and batch processing, then converted to model-readable descriptions. The library is built from real survey data, cleaned, translated, and unified into a common structured representation. To preserve inter-attribute correlations (e.g., income, consumption, assets, debts co-vary), the team uses conditional joint sampling under constraints such as age, household income, and total expenditure. Statistical calibration on selected population distributions (household size, debt, housing tenure, financial assets) is applied when scaling to 500M personas. Principal Component Analysis (PCA) on 50 data shards shows the first principal component explains 32.62% of variance on average, with stable attribute contributions across shards.

Persona Quality Validation

Quality is assessed via the MatrAIx five-dimensional framework (accuracy, unsupported inference, coverage, internal consistency, overall fidelity) using 100 processed real survey personas evaluated by 6 human raters (3000 ratings total). The mean human score is 4.570/5, with 87.8% ratings at 4 or 5. Accuracy: 4.685/5; internal consistency: 4.670/5; coverage: 4.985/5 (591/600 coverage ratings are perfect). Automated evaluation with DeepSeek and GLM shows 94.0% of model-human pairwise comparisons within 1 point; re-evaluation with DeepSeek V4 Pro and GLM 5.3 yields 94.3% and 94.0% within-1-point agreement. Injected errors (fact modification, unsupported info addition, field omission, internal contradiction) are detected by GLM with 95.8%-100% of valid pairs showing at least 1-point score drop on the relevant dimension.

Model: YN-mini Decision Model Family

Instead of autoregressive token generation, S²-MAS frames decisions as structured multiple-choice: given household state, a question, and candidate actions, the model scores each action and outputs a probability distribution via softmax. The architecture uses a pretrained backbone (ModernBERT or Qwen) plus a trainable scoring head. The team systematically compares six training routes across two backbone families:

ModernBERT encoder SFT (supervised fine-tuning on survey QA)

ModernBERT encoder SFT→CE (continued cross-entropy training)

ModernBERT encoder SFT→RLCD (cross-entropy + policy gradient with probability-based reward)

Qwen 4B frozen backbone + decision head

Qwen 4B LoRA + decision head

Qwen 9B LoRA + decision head

Ten training variants are released, spanning 421M, 322M, 4B, and 9B parameter sizes. The flagship YN-mini uses the Laya decision architecture (395M ModernBERT-large backbone + scoring head), total ~421M parameters, trained via two-stage SFT and native RLCD. Its MLX FP16 version is used for the 12M-household simulation.

In-Scenario Simulation Benchmark

On a held-out test set of 240,458 single-round consumer finance choice questions (119 question versions, 8,275 respondent household groups), YN-mini achieves the highest accuracy among 14 evaluated models: 74.55%, surpassing Jev 1.13 by 9.30 percentage points, GLM-5.3-Flash (320B params) by 5.68 points, and GPT-6-Astra by 5.39 points on a fixed stratified sample (500 groups, 14,441 records). Macro-average F1 (38.37%) also exceeds GLM-5.3-Flash, Jev 1.13, and DeepSeek-V4.1-Flash. The authors emphasize this advantage is specific to in-scenario one-shot simulation, not general reasoning.

Throughput

On a single A100 GPU, YN-mini reaches 37,000 tokens/s, 22.72× the throughput of the Jev API, turning "unaffordable" inference into "runs fast" for millions of household decisions.

Engine: Genese Simulation Engine

Genese combines microservice architecture with ECS (Entity-Component-System) acceleration to balance decoupling, transactional consistency, and speed. Key mechanisms:

Entity-component array batch updates: household state stored in structured arrays; same-type financial updates vectorized in bulk, not per-household loops.

Atomic commit of state, event log, and command results within a single transaction boundary, preventing "account changed but record missing" drift.

Snapshot rollback recovery: any partition step failure triggers automatic rollback to round-start snapshot.

Multi-partition asynchronous pipeline: data preparation, model inference, and ledger posting overlap across batches, eliminating GPU idle wait.

Pluggable application rules: consumption, financing, repayment rules plug via unified state, decision, and execution interfaces, enabling flexible scenario extension.

The monthly household ledger uses cents as the minimal currency unit, settling income, borrowing, consumption, and due repayments in fixed accounting order, with every funding gap independently logged.

Engine Benchmarks

On a controlled benchmark with 1,000 agents on a single T4 GPU, Genese completes a round in 6.17s, ~17% faster than OASIS and ~20% faster than OneSim. With a fixed decision backend (pure engine scheduling and bookkeeping), Genese runs 2.12–4.10× faster than AgentSociety, OneSim, OASIS, and MiroFish. In the 12M-household production run, tasks are split into 2,400 work units pipelined across 14 GPUs, achieving ~1,672 household decisions per second end-to-end (~886K tokens/s).

Large-Scale Experiment: 12M Households, Paired Conditions

From the persona library, 12M distinct synthetic households are sampled. Each faces two paired conditions: "¥20K consumption target + ¥20K credit limit" vs "¥30K target + ¥30K limit". Choices: defer purchase, use savings plus half borrowing, or full installment (5% APR, 24 months). The monthly ledger then projects 12-month financial trajectories. Scale: 12M households, 24M paired decisions, 288M monthly ledger states. Hardware: 14 GPUs. End-to-end time: 3h 59m 17s (14,356.98s). This exceeds prior LLM-driven multi-agent simulations (typically tens of thousands to millions of agents) per Table S9.

Heterogeneous Effects Revealed by Scale

Aggregate means: raising target/limit from ¥20K to ¥30K increases annual consumption by ¥2,887.18/household, reduces year-end cash by ~¥929, increases year-end debt by ~¥2,072, and raises simulated default rate by ~1.67 percentage points. The value of 10M+ scale is the ability to slice these averages across fine-grained subgroups while maintaining statistical stability:

Age divergence: 25–34 decision-makers show largest consumption increase (¥9,161.69 avg); 55–64 increase only ~¥13; all 2.364M households aged 65+ choose deferral in both conditions.

Education & purchase conversion: After standardizing age and income, college-educated households increase consumption by ~¥3,821 vs ~¥398 for primary-or-less. Ledger decomposition shows the gap comes from "deferral-to-purchase" conversion, not higher spending per buyer.

Asset-liability status: Households with initial debt of ¥2M–5M convert nearly all new consumption into year-end debt; zero-initial-debt households absorb more via cash drawdown and scheduled repayment.

Such granular structural heterogeneity (42 age×income grids further crossed by education, employment, assets, debts) requires millions of agents per cell to remain statistically stable.

From Scenario Simulation to Household Consumption Prediction

The team tests whether YN models can predict actual household consumption given income, cash, and policy changes. Using 1,060 real households, they compare YN-Lite-4B and YN-master-9B against a fully rational HANK-style optimization benchmark (solving optimal consumption/savings under budget constraints). YN models assign probabilities to 33 feasible consumption options each month, execute the highest-probability option, and update remaining cash. The policy scenario is China's 2018 personal income tax reform (monthly basic deduction raised from ¥3,500 to ¥5,000, rate brackets adjusted), effective Oct 1, 2018. Simulation starts from September with known policy arrangement; tax-cut income added to budget from October; predicted Sep–Dec consumption plus common first-eight-months base yields annual prediction.

Prediction Results

Per-household Mean Absolute Error (MAE): YN-master-9B achieves ¥51,312, 3.42% lower than rational benchmark; RMSE 3.15% lower. Improvement driven by high-income group (MAE from ¥90,874 to ¥83,282). Group-level bias (low/medium/high income): YN-Lite-4B reduces log-error fit distance by 14.26% vs rational benchmark. Results weighted by household survey weights; actual mean annual consumption ¥99,300.73 (2018 nominal yuan).

Contamination Check

To test if models merely memorized 2018 outcomes, the year is changed to 2031 while keeping policy description. Across 12 rephrased/option-order variants per model, highest-probability choices remain identical: 9B still selects "increase" for annual and December consumption direction, "uncertain" for magnitude; 4B selects "uncertain" for all three. Year change shifts some option probabilities but not final choices.

Open Resources and Democratizing Large-Scale Simulation

The team releases: YN decision models (10 training variants across 6 routes), full training and evaluation code, research plotting and result reproduction scripts, 2M independent synthetic household personas, and 32K unlabeled task inputs. Model weights on Hugging Face (ya9919/s-mas-models-sharp), datasets (ya9919/s-mas-dataset-resources), code on GitHub (KataDavidXD/s2-mas), and author blog (huggingface.co/blog/ya9919/s2-mas). The paper (arXiv:2610.27746) includes supplementary methods and tables. The goal is to lower the entry barrier for society-scale simulation, enabling resource-constrained teams to run larger populations, finer slices, and more counterfactual comparisons, and to foster a community accumulating reusable experimental methods and research experience toward policy rehearsal, marketing forecasting, transaction prediction, and beyond.