Research map / Large language models / Agents, code and math
MF-LLM: Simulating Population Decision Dynamics via a Mean-Field Large Language Model Framework
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MF-LLM, which alternates an LLM mean-field model (population signal) and LLM policy model for social simulation, and IB-Tune, an information-bottleneck fine-tuning method, evaluated on Weibo rumor/event response data.
Problem
LLM-based social simulations are prompt-centric and fail to quantitatively match real-world population decision dynamics, while SFT ignores agent-to-agent interaction dynamics and full pairwise interaction modeling does not scale.
Approach
Replace agent-to-agent interactions with agent-population interactions via a mean-field LLM that recursively summarizes population behavior into a textual signal, and fine-tune the signal model with an information-bottleneck objective that keeps the signal predictive of the next actions while compressing history.
Key results
- MF-LLM IB-Tune on Qwen2-1.5B reduces KL divergence to real distributions by 47.0% relative to the State baseline (0.512 vs 0.966).
- Without fine-tuning, MF-LLM improves KL over State by 60.5% on GPT-4o-mini and 34.2% on DeepSeek-R1-32B.
- Removing the mean-field module raises KL divergence by 118.7% relative to MF-LLM in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use mean-field population signals instead of the State baseline | MF-LLM with GPT-4o-mini · Weibo | KL divergence: 4.172 -> 1.647 | Helped |
| Use mean-field population signals instead of the State baseline | MF-LLM with DeepSeek-R1-32B · Weibo | KL divergence: 1.946 -> 1.280 | Helped |
| Use mean-field population signals instead of the State baseline | MF-LLM with Qwen2-7B-Instruct · Weibo | KL divergence: 1.153 -> 1.010 | Helped |
| Use mean-field population signals instead of the State baseline | MF-LLM with Qwen2-1.5B-Instruct · Weibo | KL divergence: 0.966 -> 0.645 | Helped |
| Fine-tune mean-field and policy models with IB-Tune | MF-LLM with Qwen2-1.5B-Instruct · Weibo | KL divergence: 0.645 -> 0.512 | Helped |
| Add IB-Tune fine-tuning to the mean-field model (ablation) | MF-LLM · Weibo ablation | KL divergence: 0.4813 -> 0.5312 | Helped |
| Add IB-Tune fine-tuning to the policy model (ablation) | MF-LLM · Weibo ablation | KL divergence: 0.4813 -> 0.5882 | Helped |
| Add IB-Tune fine-tuning to both models (ablation) | MF-LLM · Weibo ablation | KL divergence: 0.4813 -> 0.6166 | Helped |
| Add the mean-field module and IB-Tune instead of using SFT without mean field (ablation) | MF-LLM · Weibo ablation | KL divergence: 0.4813 -> 0.7531 | Helped |
| Add the mean-field module and IB-Tune instead of using a pretrained policy without mean field (ablation) | MF-LLM · Weibo ablation | KL divergence: 0.4813 -> 1.0526 | Helped |
| Add the mean-field module and IB-Tune instead of using SFT without mean field (ablation) | MF-LLM · Weibo ablation | NLL loss: 2.4217 -> 2.9029 | Hurt |
Surprising findings
- Smaller backbones (Qwen2-1.5B) outperform larger models (GPT-4o-mini, DeepSeek-R1-32B) on population dynamics fidelity, attributed to homogeneous responses from larger models.
- SFT achieves the lowest NLL loss yet has substantially worse KL divergence and trajectory fidelity than MF-LLM.
Limitations
- Small deviations in early steps compound over long horizons, causing distributional drift.
- Does not support real-time detection or injection of exogenous signals; these are injected manually.
- Evaluation relies on GPT-4o-mini as judge for semantic dimensions.
Experimental setup
Datasets: Weibo corpus
Benchmarks: Weibo corpus
Metrics: KL divergence Wasserstein distance DTW distance NLL loss Macro F1 Micro F1
Baselines: State Recent Popular SFT OASIS AgentSociety
Builds on: mean field theory information bottleneck SFT
Closest papers by meaning
- Position: LLM Social Simulations Are a Promising Research Method (ICML 2025 · similarity 0.89)
- TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets (NeurIPS 2025 · similarity 0.87)
- LLM Active Alignment: A Nash Equilibrium Perspective (NeurIPS 2026 · similarity 0.87)
- Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces (NeurIPS 2026 · similarity 0.87)
- Bias Amplification in Language Model Evolution: An Iterated Learning Perspective (NeurIPS 2024 · similarity 0.86)
- Investigating Prosocial Behavior Theory in LLM Agents Under Policy-Induced Inequities (AAAI 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.