M2Sim

Macro dynamics · Micro states

Coupling Macro Dynamics and Micro States for Long-Horizon Social Simulation

Macro–micro coupled social simulation with evolving individual states and collective dynamics.

  • Yunyao Zhang1
  • Yihao Ai2
  • Zuocheng Ying1
  • Qirui Mi3
  • Junqing Yu1
  • Wei Yang1
  • Zikai Song1†

1 Huazhong University of Science and Technology, Wuhan, China

2 National University of Singapore, Singapore

3 Institute of Automation, Chinese Academy of Sciences, Beijing, China

† Corresponding author: skyesong@hust.edu.cn

Paper Code BibTeX
Framework overviewMacro ↔ Micro
M²Sim couples a macro Event Transformer and its K-step rollout consistency objective with a micro state-conditioned policy LLM and rollout-based action reselection.
M2Sim overview. At the macro level, an Event Transformer updates the distributional mean field with K-step rollout consistency. At the micro level, a state-conditioned policy LLM generates candidate actions and uses their predicted future impact for action reselection. View full-resolution figure ↗

01 / Research context

Abstract

Social network simulation aims to model collective opinion dynamics in large populations, but existing LLM-based simulators primarily focus on aggregate dynamics and largely ignore individual internal states. This limits their ability to capture opinion reversals driven by gradual individual shifts and makes them unreliable over long-horizon simulations.

Unlike existing methods that collapse dynamics into macro-only updates, we propose a social simulation framework, M2Sim, which tightly couples macro-level collective dynamics with micro-level individual states.

We explicitly model micro-level opinion states with a state transition mechanism, coupling Macro-level collective dynamics with Micro-level states. This allows the opinion-state distribution to evolve gradually, rather than triggering instant reactions, enabling the simulator to distinguish states close to or far from switching, capture opinion reversals, and maintain accuracy over long-horizon simulations.

Across real-world events, our M2Sim supports stable simulation of long-horizon social events with up to 40,000 interactions, while reducing long-horizon KL divergence by 75.3% (1.2490 → 0.3089) and reversal KL by 66.9% (1.6425 → 0.5434), significantly mitigating the drift observed in MF-LLM.

02 / Contributions

Highlights

M2Sim addresses three central challenges in social simulation: connecting individual and collective dynamics, sustaining long-horizon evolution, and capturing opinion reversals.

  1. Micro–Macro Coupling

    M2Sim couples macro-level collective dynamics with micro-level opinion states through an explicit state-transition mechanism. Agents act on both their individual states and the evolving mean field, connecting individual behavior with collective opinion evolution.

  2. Long-Horizon Dynamics

    Gradual changes in opinion require modeling beyond instantaneous reactions. A learned state-transition model and multi-step rollout consistency help preserve collective trajectories over long horizons, addressing accumulated drift, variance damping, and majority lock-in.

  3. Opinion Reversals

    Capturing a reversal requires tracking when opinions change and how far they shift. By conditioning actions on evolving opinion states and evaluating their multi-step impacts, M2Sim supports action reselection in response to exogenous signals, addressing the inertia that can obscure turning points.

03 / Inside the framework

Method

Dual channels
The text synopsis rt summarizes interactions and signals; the distribution mt tracks opinion-state proportions. Both condition micro agents alongside their states, personalities, and current signals.
Recursive feedback
The Event Transformer updates participant states and the distribution. Realized actions update the text synopsis through the summarizer; the resulting pair becomes the next step’s macro context.
Training-only rollout
Self-conditioned predictions support transition consistency and candidate-policy scoring. Discounted KL costs guide soft weighting and action reselection, with auxiliary text supervision. Observed future distributions supervise training; agents never receive the real future during simulation.

04 / Experimental evidence

Results

Trajectory alignment, main comparisons, and component ablations across short-horizon, long-horizon, and reversal simulations.

Long-horizon action and opinion-state trajectories, and an opinion reversal event, comparing real data with M²Sim, MF-LLM, Social Retrieval, and Direct LLM.
Trajectory alignment of collective actions and state distributions. Long-horizon and reversal trajectories compare M²Sim with real data and three baselines. View full-resolution figure ↗

Main experiments

Comparison across the three simulation settings reported in the paper.

Main experiment results
Method KL Div. ↓ Wass. Dist. ↓ DTW ↓ Macro F1 ↑ Micro F1 ↑ NLL Loss ↓
Short-Horizon · Default Steps
Direct LLM 0.1101 0.1661 0.1578 0.5805 0.6953 4.0674
Social Retrieval 0.1051 0.1578 0.1500 0.5829 0.6905 3.9532
MF-LLM 0.0492 0.1062 0.0944 0.5861 0.6975 3.9336
M²Sim (Ours) 0.0453 0.1006 0.0995 0.5897 0.7082 3.9156
Long-Horizon · Full Steps
Direct LLM 1.8300 0.3746 0.4003 0.3922 0.5621 4.6141
Social Retrieval 1.6554 0.3515 0.3816 0.4014 0.5962 4.5294
MF-LLM 1.2490 0.3251 0.2886 0.4643 0.6153 3.9198
M²Sim (Ours) 0.3089 0.1773 0.1666 0.4805 0.6393 3.7655
Reversal · Full Steps
Direct LLM 5.4883 0.4535 0.4377 0.3531 0.4837 4.6469
Social Retrieval 3.5539 0.4432 0.3978 0.3690 0.5017 4.2002
MF-LLM 1.6425 0.2763 0.2425 0.4158 0.5721 3.8841
M²Sim (Ours) 0.5434 0.2127 0.1986 0.4533 0.6065 3.8384

Mean values from Table 1. ↓ Lower is better; ↑ higher is better. Bold marks the best value within each setting. Relative improvement percentages are omitted for readability.

Ablation study

Removing LCT-State, LCT-Policy, or Sampling from the full framework.

Component ablation results
Method KL Div. ↓ Wass. Dist. ↓ DTW ↓ Macro F1 ↑ Micro F1 ↑ NLL Loss ↓
Short-Horizon · Default Steps
M²Sim (Ours) 0.0453 0.1006 0.0995 0.5897 0.7082 3.9156
w/o LCT-State 0.0449 0.1120 0.0802 0.5825 0.6828 4.0064
w/o LCT-Policy 0.0491 0.1051 0.0856 0.5809 0.6962 3.9403
w/o Sampling 0.0649 0.1115 0.1081 0.5925 0.7014 3.9206
Long-Horizon · Full Steps
M²Sim (Ours) 0.3089 0.1773 0.1666 0.4805 0.6393 3.7655
w/o LCT-State 0.8770 0.2123 0.2104 0.4725 0.6187 3.9414
w/o LCT-Policy 0.5872 0.2028 0.1730 0.4314 0.5765 3.9738
w/o Sampling 0.3536 0.1885 0.1710 0.4482 0.5774 3.7782
Reversal · Full Steps
M²Sim (Ours) 0.5434 0.2127 0.1986 0.4533 0.6065 3.8384
w/o LCT-State 1.1748 0.2536 0.2319 0.4649 0.6232 4.0610
w/o LCT-Policy 0.9359 0.2396 0.1808 0.4309 0.5592 4.2301
w/o Sampling 0.7634 0.2281 0.1897 0.4214 0.5721 4.0404

Mean values from Table 2. ↓ Lower is better; ↑ higher is better. Bold marks the best value within each setting. Relative change percentages are omitted; metric values retain the paper’s precision.

05 / Work with M2Sim

Usage

Clone the repository, install its dependencies with Poetry, and create a local environment configuration.

git clone https://github.com/AI4SS/M2Sim.git
cd M2Sim
poetry install
cp .env.example .env

Configure .env for your environment and supply processed event JSON data, trained policy and mean-field model paths, and an Event Transformer checkpoint. In the configured Python environment, the single-event entrypoint is:

python main/script/run_m2sim.py \
  --file_name /path/to/event.json \
  --simulation_start 50 \
  --model 1.5B \
  --batch_size 16 \
  --st_model_path /path/to/event_transformer_best.pt

Replace the placeholder paths with your data and checkpoint. Set --policy_model_path and --mf_model_path when supplying trained model directories explicitly. See the repository README for full setup, data preparation, and evaluation instructions.

Reference

Citation

If this work is useful to your research, please cite the paper.

Download .bib 
@misc{zhang2026couplingmacrodynamicsmicro,
  title={Coupling Macro Dynamics and Micro States for Long-Horizon Social Simulation},
  author={Yunyao Zhang and Yihao Ai and Zuocheng Ying and Qirui Mi and Junqing Yu and Wei Yang and Zikai Song},
  year={2026},
  eprint={2604.05516},
  archivePrefix={arXiv},
  primaryClass={cs.SI},
  url={https://arxiv.org/abs/2604.05516},
}