OpenSim

Open populations · Observable communication

Rethinking Social Simulation in Open-Population Systems: From Persistent Agents to Communication Processes

Who participates, what they communicate, and how communication shapes the social process.

¹ Huazhong University of Science and Technology, Wuhan, China

² National University of Singapore, Singapore

^* Equal contribution. · ^† Corresponding author: skyesong@hust.edu.cn

Paper Code Video BibTeX

OpenSim models a platform process under declared external inputs. Participants change across activations while observable communication accumulates into a continuous event history. — **Event history continues as participants change.** Full-resolution figure ↗

A communication-centered perspectiveProject video / 01:34

A 94-second concept demonstration. The users, posts, values, and trajectories shown are illustrative examples, not experimental results. Download video (MP4) ↗

01 / Research context

Abstract

Large language models (LLMs) have enabled social simulation with persistent agents, where fixed computational individuals maintain internal states and interact over time.

However, large-scale social networks are inherently open-population: the participating population changes over time, users enter and leave dynamically, and only a sparse subset becomes observable through communication at any given time. In such systems, persistent account identities do not provide continuously observable individual-state trajectories, making recursively maintained psychological states insufficiently grounded in empirical observations.

We therefore shift the simulation target from persistent individual trajectories to the observable evolution of social communication, modeling who participates, what they communicate, and how these communications shape subsequent system dynamics. We instantiate this formulation with participant activation from a global user pool, collective state transition, and an LLM policy for communication generation, while explicitly separating environmental inputs from endogenous system evolution.

Across five Qwen backbones, our framework reduces average KL divergence by 71.2% over Direct LLM and 35.0% over MF-LLM, with boundary and counterfactual experiments further validating the proposed modeling boundaries. Our results support communication-centered simulation as a principled formulation for long-horizon open-population social systems.

02 / Contributions

Highlights

A shift in what the simulator carries forward, what it predicts, and where its responsibility ends.

Communication carries continuity
Observable communication history carries the recursive process. Participants realize communicative events without requiring the simulator to recover their unobserved psychological trajectories.
Participation is endogenous
The framework models participation and communication conditional on declared environmental inputs, separating the platform process from external events it is not tasked with inventing.
Learn from future dynamics
Future-constrained learning connects plausible communication with later collective evolution. Generation and selection use available context and model predictions, without the realized future participant sequence.

03 / Inside the framework

Method

OpenSim forecasts communication volume, activates participants, infers activation-local states, generates and selects communication, and updates the observable platform history. — **Overview of OpenSim.** OpenSim recursively models participation, mean-field state transition, and future-constrained communication generation, with realized communication updating the collective state and platform history for subsequent steps. Full-resolution figure ↗

Open-population participation: Given prior platform history and declared environmental inputs, OpenSim forecasts communication volume and activates participants from a candidate pool. Participation is step-specific: users may enter, leave, and reappear, while history carries continuity. The manuscript distinguishes event-specific and external participant pools.
Collective state and communication: A state-transition model predicts attitude composition and activation-local representations. A policy uses these predictions, participant features, and history to generate actions and content. Selected communications update the realized mean field and textual synopsis. Latent representations are not maintained through inactive periods.
Future-constrained learning: A response-sensitive critic learns from communication supervision, trajectory discrepancy, and matched versus masked or mismatched responses. During simulation, candidates are ranked against model-predicted future trajectories. Observed future distributions provide training supervision only.

04 / Experimental evidence

Results

Main comparisons, component ablations, and counterfactual interventions reported in the manuscript.

Main experiments

Across five backbones, the event-specific-pool OpenSim row reduces mean KL by 71.2% relative to Direct LLM* and 35.0% relative to MF-LLM*. These percentages compare the means across backbones; they do not describe OpenSim†.

Comparison conditions. * Baselines receive the observed future participant sequence. OpenSim uses an event-specific pool derived from that event’s records; OpenSim† uses an external participant pool. Event context and stepwise counts are held fixed in the pool comparison.

Expand a backbone to inspect all six metrics. ↓ Lower is better; ↑ higher is better. Bold marks the best value within that backbone.

Qwen2-1.5B5 settings · 6 metrics

Qwen2-1.5B — main results reported in the manuscript
Method / setting	KL ↓	Wass. ↓	DTW ↓	Macro-F1 ↑	Micro-F1 ↑	NLL ↓
Direct LLM^*	2.4090	0.3336	0.3055	0.3441	0.5686	4.6838
Social Retrieval^*	1.6678	0.2568	0.2063	0.3636	0.5526	5.1355
MF-LLM^*	1.2494	0.2246	0.1773	0.4311	0.6295	4.6060
OpenSim^†	2.0773	0.2905	0.2256	0.4209	0.6199	4.1414
OpenSim	1.1352	0.1922	0.1261	0.4444	0.6616	4.3377

Qwen2.5-1.5B5 settings · 6 metrics

Qwen2.5-1.5B — main results reported in the manuscript
Method / setting	KL ↓	Wass. ↓	DTW ↓	Macro-F1 ↑	Micro-F1 ↑	NLL ↓
Direct LLM^*	3.3485	0.2844	0.3384	0.3730	0.4651	6.3707
Social Retrieval^*	2.4112	0.2243	0.2695	0.4569	0.5566	6.2568
MF-LLM^*	1.4739	0.1941	0.1396	0.5408	0.5881	5.1429
OpenSim^†	1.5739	0.2142	0.1934	0.5001	0.6008	4.4138
OpenSim	1.0667	0.1724	0.1252	0.6097	0.6252	4.5980

Qwen2.5-3B5 settings · 6 metrics

Qwen2.5-3B — main results reported in the manuscript
Method / setting	KL ↓	Wass. ↓	DTW ↓	Macro-F1 ↑	Micro-F1 ↑	NLL ↓
Direct LLM^*	3.8755	0.3603	0.3771	0.3138	0.4561	13.9996
Social Retrieval^*	1.8466	0.2283	0.2591	0.3817	0.5129	14.0303
MF-LLM^*	1.5061	0.1546	0.2311	0.4598	0.6222	4.5446
OpenSim^†	1.5030	0.1495	0.1824	0.4652	0.6545	4.4980
OpenSim	1.0048	0.1273	0.1369	0.5058	0.7278	4.5356

Qwen2.5-7B5 settings · 6 metrics

Qwen2.5-7B — main results reported in the manuscript
Method / setting	KL ↓	Wass. ↓	DTW ↓	Macro-F1 ↑	Micro-F1 ↑	NLL ↓
Direct LLM^*	3.0914	0.3107	0.3097	0.3471	0.5581	13.0946
Social Retrieval^*	2.5142	0.2981	0.2882	0.3701	0.5821	12.8823
MF-LLM^*	1.6049	0.2234	0.2564	0.4380	0.6117	4.6823
OpenSim^†	1.4294	0.2289	0.2509	0.4455	0.6340	4.9849
OpenSim	0.8509	0.1759	0.1572	0.4855	0.6752	5.0921

Qwen3-1.7B5 settings · 6 metrics

Qwen3-1.7B — main results reported in the manuscript
Method / setting	KL ↓	Wass. ↓	DTW ↓	Macro-F1 ↑	Micro-F1 ↑	NLL ↓
Direct LLM^*	5.1480	0.4279	0.4191	0.3314	0.5236	25.7470
Social Retrieval^*	4.7743	0.3453	0.2210	0.3478	0.5451	25.4819
MF-LLM^*	2.0902	0.2549	0.1715	0.5095	0.6518	6.3621
OpenSim^†	1.3115	0.2377	0.1583	0.4865	0.6825	4.4081
OpenSim	1.0929	0.1780	0.1098	0.5445	0.7610	4.4414

All 150 main-result metric values are retained from manuscript Table 1; parenthetical improvement percentages are omitted.

Manuscript case study comparing communication attitude trajectories under baseline and communication intervention conditions. — **Communication intervention dynamics.** The manuscript illustrates how changes in communication reshape subsequent collective trajectories. Full-resolution figure ↗

Component ablations

Removing state-transition supervision or the future-constrained policy objective, with participation modeling and the remaining pipeline unchanged.

Component ablation plots comparing full OpenSim, removal of the state-transition objective, and removal of the future constraint across three backbones. — **Component ablations across backbones.** Full OpenSim improves trajectory metrics in the reported comparisons; removing future constraints can improve NLL for some backbones. Full-resolution figure ↗

Expand a backbone for the complete ablation values. Bold marks the best value within that backbone.

Qwen2-1.5B3 settings · 6 metrics

Qwen2-1.5B — ablation results reported in the manuscript
Method / setting	KL ↓	Wass. ↓	DTW ↓	Macro-F1 ↑	Micro-F1 ↑	NLL ↓
OpenSim	1.1352	0.1922	0.1261	0.4444	0.6616	4.3377
w/o ST Objective	3.1716	0.3074	0.2644	0.4087	0.5079	4.4791
w/o Future Constraint	1.6934	0.2410	0.1914	0.4331	0.5597	4.4315

Qwen2.5-1.5B3 settings · 6 metrics

Qwen2.5-1.5B — ablation results reported in the manuscript
Method / setting	KL ↓	Wass. ↓	DTW ↓	Macro-F1 ↑	Micro-F1 ↑	NLL ↓
OpenSim	1.0667	0.1724	0.1252	0.6097	0.6252	4.5980
w/o ST Objective	2.9886	0.2725	0.2494	0.4222	0.5299	4.4942
w/o Future Constraint	1.5839	0.1940	0.1864	0.5369	0.5646	4.3211

Qwen3-1.7B3 settings · 6 metrics

Qwen3-1.7B — ablation results reported in the manuscript
Method / setting	KL ↓	Wass. ↓	DTW ↓	Macro-F1 ↑	Micro-F1 ↑	NLL ↓
OpenSim	1.0929	0.1780	0.1098	0.5445	0.7610	4.4414
w/o ST Objective	2.8526	0.2775	0.2412	0.4782	0.5776	4.5111
w/o Future Constraint	1.7857	0.2320	0.1919	0.4413	0.6587	4.0974

All 54 metric values are retained from the manuscript’s component-ablation table. w/o ST Objective removes state-transition supervision; w/o Future Constraint removes the future-constrained policy objective.

Results are reported from the manuscript. See the released dataset and evaluation disclosure for record provenance and the reference-text matching protocol.

05 / Work with OpenSim

Usage

The release requires Python 3.11+, CUDA, and Git LFS. Its default backbone is Qwen/Qwen3-1.7B; model weights are not included.

git lfs install
git clone https://github.com/AI4SS/OpenSim.git
cd OpenSim
git lfs pull
python -m venv .venv
source .venv/bin/activate
python -m pip install -e .
python scripts/run.py --dry-run

Training, simulation, and evaluation use the same workflow configuration. Simulation requires trained checkpoints and a passed validation gate.

python scripts/train.py --base-model Qwen/Qwen3-1.7B --output-dir outputs
python scripts/simulate.py --base-model Qwen/Qwen3-1.7B --output-dir outputs --resume

The released data mix original-source-derived, generated or rewritten, and unresolved-provenance records. Its micro-level evaluator samples reference records with replacement from an event-level pool; those metrics measure consistency with sampled reference texts. This differs from the manuscript’s participant-sampling protocol. See the dataset and evaluation disclosure and README for the full setup, checkpoint requirements, and evaluation instructions.

Reference

Citation

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

Download .bib

@misc{zhang2026rethinkingsocialsimulationopenpopulation,
  title={Rethinking Social Simulation in Open-Population Systems: From Persistent Agents to Communication Processes},
  author={Yunyao Zhang and Yihao Ai and Boqi Xu and Zuocheng Ying and Junqing Yu and Zikai Song},
  year={2026},
  eprint={2610.37687},
  archivePrefix={arXiv},
  primaryClass={cs.MA},
  url={https://arxiv.org/abs/2610.37687},
}

OpenSim project video