PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yunyao Zhang, Yihao Ai, Zuocheng Ying, Qirui Mi, Junqing Yu, Wei Yang, Zikai Song

arXiv:2604.05516v2cs.SI

TL;DR

Existing social simulators often emphasize aggregate dynamics while omitting individual latent states, limiting their treatment of gradual opinion change and reversals. MF-MDP couples per-agent state transitions with a macro mean-field model and improves long-horizon stability and reversal tracking across real-world events. The paper also notes evaluation noise from LLM-generated semantic labels and broader risks of misuse.

  • Problem

    Existing simulation approaches struggle to represent evolving beliefs, delayed commitment, and coupled long-horizon opinion dynamics because they rely on static rules or omit micro-agent states.

  • Method

    MF-MDP couples macro-level collective dynamics with micro-level agent states through a Mean-Field Markov Decision Process and explicit state transitions.

  • Results

    MF-MDP consistently outperforms existing methods across real-world events, reducing long-horizon KL divergence by 75.3% and reversal KL by 66.9% while supporting up to 40,000 interactions versus ∼300.

  • Takeaways & Limitations

    The framework supports stable long-horizon social simulation while better tracking opinion reversals and mitigating drift in prior mean-field simulators.

  • Takeaways & Limitations

    Some annotation noise may remain because semantic labels are produced by an LLM annotator, despite evaluation with explicit dimensions and quantitative trajectory metrics.

Abstract

from arXiv · show

Social network simulation aims to model collective opinion dynamics in large populations, but existing LLM-based simulators mainly focus on aggregate dynamics while largely ignoring individual internal states. This limits their ability to capture opinion reversals driven by gradual individual shifts and makes them unreliable in long-horizon simulations. We propose MF-MDP, a social simulation framework that tightly couples macro-level collective dynamics with micro-level individual states. MF-MDP explicitly models per-agent latent opinion states with a state transition mechanism, combining individual Markov Decision Processes at the micro level with a mean-field collective framework at the macro level. This allows individual behaviors to change internal states gradually rather than trigger instant reactions, enabling the simulator to distinguish agents that are close to switching from those that are far from switching, capture opinion reversals, and maintain accuracy over long horizons. Across real-world events, MF-MDP supports stable simulation of long-horizon social processes with up to 40,000 interactions, compared with about 300 in the baseline MF-LLM, while reducing long-horizon KL divergence by 75.3% (1.2490 to 0.3089) and reversal KL by 66.9% (1.6425 to 0.5434), significantly mitigating the drift observed in MF-LLM. Code is available at github.com/AI4SS/MF-MDP.

1. Introduction

Social simulation must capture tightly coupled macro dynamics and evolving individual states, because existing approaches struggle with delayed commitment, long-horizon degradation, and opinion reversals. MF-MDP addresses these challenges with a coupled mean-field and state-aware framework and shows stronger long-horizon performance.

  • Existing mechanistic, statistical, and agent-based models rely on static parameters or handcrafted rules, limiting their representation of evolving beliefs and delayed commitment at scale.
  • LLM-based simulators generate rich behaviors but often treat actions as instantaneous reactions without explicit per-agent latent states, weakening reversal and long-horizon dynamics.
  • Micro-macro decoupling, accumulated rollout errors, and unresolved opinion reversals are identified as three central challenges for social simulation.
  • MF-MDP formulates social simulation as a Mean-Field Markov Decision Process coupling macro signals with state-conditioned decisions based on micro states.
  • 75.3% lower long-horizon KL divergence and 66.9% lower reversal KL accompany stable simulations of up to 40,000 interactions versus ∼300 for MF-LLM.

2. Background and Motivation

The framework models collective social dynamics through a mean field while retaining agent-level states, exposing how state-ignorant aggregation can systematically reinforce an existing majority and miss reversals.

  • Background and motivation: State-ignorant dynamics can induce self-reinforcing rollouts that miss opinion reversals.The formal analysis identifies this failure as a consequence of omitting micro-level agent states.
  • Framework: MF-MDP couples macro-level mean-field dynamics with micro-level agent states, actions, personalities, and exogenous signals.The macro state captures collective behavioral trends, while agent decisions are conditioned on mean-field signals and individual information.
  • Framework: The simulation maintains both a textual synopsis and a distributional mean field to represent recent collective trajectories and public mood.Together, these macro representations summarize short-term responses and longer-term collective dynamics.
  • Background and motivation: The existing baseline iteratively samples agent actions from public mood and personality, then summarizes the next public mood from actions and the current mood.This update process continues until convergence or a termination condition is reached.
  • Background and motivation: δm(s*) ≥ η − ϵ ≥ 0 after the last exogenous signal when η ≥ ϵ, formalizing nonnegative drift toward the current majority state.The proposition characterizes the self-strengthening behavior of state-ignore dynamics under mild assumptions.
  • Background and motivation: A small aggregation error makes one-step majority drift nonnegative when a clear public-mood majority exists, making reversals increasingly rare.The analysis does not rule out reversals, but indicates that state-ignorant dynamics systematically reinforce the current majority after exogenous signals end.

3. Methodology

MF-MDP couples distributional and textual macro states with micro-level agent states, policies, and state transitions. Its training and rollout procedures target stable long-horizon dynamics, delayed commitment, and realistic opinion reversals.

  • MF-MDP Framework: MF-MDP formulates social propagation as a mean-field Markov decision process coupling macro-level state distributions with micro-level agent dynamics.The mean field represents collective states, while the MDP represents individual state-conditioned decisions.
  • MF-MDP Framework: The macro state uses synchronized textual and distributional channels: a trajectory synopsis and a distribution over latent agent states.The distributional synopsis can be an empirical histogram of agent states, with related states merged into coarse buckets when needed.
  • Recursive Update: At each step, a state-transition model updates agent states and the distributional mean field before the policy selects actions from private states and macro signals.Agent decisions condition on internal state, textual synopsis, distributional mean field, exogenous signal, and personality.
  • Recursive Update: The simulation then updates the textual synopsis and jointly transitions the macro channels, exogenous signal, and active agent pool.The algorithm uses a warm-up phase with background trajectories before sampling simulated actions, and initializes the distributional channel from historical state frequencies.
  • MF-MDP Framework: The framework recovers macro-only two-LLM simulation as a degenerate case when internal states and the distributional mean field are ignored.This identifies the prior simulator as a special case of the broader coupled formulation.
  • Long-Horizon Consistency Training: LCT improves MF-LLM through predictive macro signals and behaviorally realistic micro-level actions conditioned on those signals.Its macro model uses a temporal Transformer with attention over past contexts, while rollout consistency encourages stable multi-step predictions and realistic delayed commitment and reversals.

4. Experiment

MF-MDP is evaluated against LLM-based social simulation baselines on short-horizon, long-horizon, and reversal settings, with ablations isolating its components. It provides modest short-horizon gains but substantially improves long-horizon consistency and reversal tracking.

  • Experimental setup: MF-MDP is compared with Direct LLM, Social Retrieval, and MF-LLM using micro- and macro-level metrics on default, full-trajectory, and reversal rollouts.The evaluation reports KL divergence, Wasserstein distance, DTW, NLL, Macro-F1, and Micro-F1.
  • Short-horizon performance: MF-LLM already achieves strong near-term distributional and classification accuracy, while MF-MDP adds only modest but consistent short-horizon improvements.The results suggest explicit state distributions and stateful dynamics contribute mainly beyond short horizons.
  • Long-horizon robustness: 75.3%: MF-MDP reduces long-horizon KL from 1.2490 to 0.3089 relative to MF-LLM, while DTW decreases from 0.2886 to 0.1666.MF-MDP remains stable as baselines degrade over full-trajectory rollouts.
  • Reversal opinion dynamics: 66.9%: MF-MDP reduces reversal KL from 1.6425 to 0.5434, while reversal DTW decreases from 0.2425 to 0.1986.Reversal Micro-F1 also increases from 0.5721 to 0.6065.
  • Ablation study: Removing LCT-State worsens long-horizon KL from 0.3089 to 0.8770 and reversal KL from 0.5434 to 1.1748, showing its importance for state-consistent evolution.The degradation grows as multi-step macro rollouts extend.
  • Ablation study: Removing Sampling alone lowers Micro-F1 from 0.6393 to 0.5774 in long-horizon simulation and from 0.6065 to 0.5721 in reversal simulation.The ablation identifies sampling-based long-horizon selection as a major contributor to F1 gains.
  • Ablation study: Under short horizons, removing LCT constraints can slightly improve distance metrics or reversal F1, but these gains trade off against coherent long-horizon macro evolution.Without LCT-State, reversal Macro-F1 rises from 0.4533 to 0.4649 and Micro-F1 from 0.6065 to 0.6232 despite worse tracking.

5. Analysis and Discussion

The analysis attributes MF-MDP’s gains to explicit macro distributions that constrain micro actions and to state coupling that allows sharp, delayed reversals. These mechanisms reduce drift in collective trajectories and help preserve turning-point timing.

  • Failure modes: MF-LLM and related macro-driven simulators can over-compress individual dynamics, treating actions as instantaneous reactions and weakening delayed commitment and reversal timing.This contributes to drift and unrealistic long-horizon collective dynamics.
  • Macro-to-micro coupling: MF-MDP predicts an explicit macro distribution and injects it into each agent’s conditioning state, creating a direct macro-to-micro constraint.Lower predicted Neutral mass produces fewer neutral-style responses and fewer repost actions, bringing action shares closer to real trajectories.
  • Long-horizon stability: Explicit state trajectories remain aligned with future distributions rather than drifting while the textual synopsis remains plausible.The state transition model is trained to match future distributions.
  • Reversal dynamics: MF-LLM often misses reversal turning points because temporally smooth synopses preserve the earlier storyline and keep micro actions inertial.The failure occurs when many agents must cross state boundaries after a delayed shift.
  • Reversal dynamics: MF-MDP can represent sharp redistribution across states and select rollouts that produce the correct non-monotonic trajectory.Agents directly observe the distribution, allowing the action mixture to pivot without requiring a gradual narrative bridge.

6. Conclusion

MF-MDP couples explicit per-agent latent opinion states with a learned state transition model and trajectory-aware action reselection. Across real-world events, it improves short-horizon fidelity, long-horizon stability, and reversal tracking relative to prior mean-field simulators.

  • Framework and contribution: MF-MDP formulates social simulation as a stateful mean-field simulator in which actions change explicit per-agent latent opinion states.A learned state transition model supplies the distributional mean field, while long-horizon trajectory agreement scores candidate actions.
  • Empirical conclusion: MF-MDP improves short-horizon fidelity, strengthens long-horizon stability, and better tracks reversals across real-world events.The framework is designed to mitigate drift in prior macro-only mean-field simulators.

Impact Statement

The paper frames MF-MDP as an analytical and simulation tool for studying opinion dynamics and testing interventions in controlled settings. It also identifies misuse risks and residual annotation noise, motivating ethical safeguards and transparency.

  • Potential benefits: MF-MDP may support understanding collective behavior and testing interventions in controlled social-network settings.The stated benefits concern analysis and simulation rather than deployment guidance for influencing individuals.
  • Risks and safeguards: The same techniques could be misused to optimize persuasion, amplify misinformation, or support manipulation at scale.The paper therefore recommends ethical review, transparency, and safeguards.
  • Evaluation limitation: LLM-produced semantic labels introduce possible annotation noise, although evaluation uses explicit discrete dimensions and quantitative trajectory metrics.Reported metrics include F1 scores and distributional distances.
  • Research context: Traditional and agent-based paradigms often rely on handcrafted rules, simplified assumptions, or fixed parameters, limiting adaptability in evolving long-horizon settings.LLM-driven simulation and mean-field methods are presented as newer alternatives for scalable social modeling.

B. Derivation of the Policy Model

The policy model uses latent-policy optimization and long-horizon rollout costs to select actions under evolving mean-field conditions. Its objective combines rollout discrepancy with text supervision while controlling variance through soft weighting over latent policy instances.

  • The policy model governs microscopic decisions through a factorized policy over the active agent set.
  • Latent-policy optimization introduces λ-conditioned policies and regularizes their distribution toward a tractable reference policy.
  • The inner optimization yields a log-sum-exp objective, inducing softmax weighting over latent policy instances.
  • The long-horizon prediction loss aggregates discounted divergence across simulation time and active agents, with auxiliary text supervision added to stabilize optimization.
  • The variance bound decreases as the selected action probability πϕ(a*|λ) increases, while dropout sampling instantiates latent policy variations in practice.

C.1. Training Curves and Hyperparameters

Training uses separate objectives for the Event Transformer and policy model, and evaluation combines micro-level semantic annotations with macro-level distributional metrics. The benchmark uses WEIBO events and compares MF-MDP with three baselines.

  • Training curves: The Event Transformer minimizes KL-based transition loss, while the policy model minimizes Ltotal = Lpred + α Ltext.Lpred captures discounted K-step rollout divergence and Ltext provides ground-truth action NLL supervision.
  • Dataset: The benchmark uses WEIBO events with temporally ordered responses and individual profiles across multiple social-event categories.
  • Baselines: Figures compare MF-MDP with MF-LLM, Social Retrieval, and Direct LLM across semantic dimensions and distributional metrics.
  • Evaluation metrics: Micro-level evaluation annotates generated and real actions across eight semantic dimensions, including sentiment, state, behavior, stance, belief, subjectivity, intent, and rumor.
  • Evaluation metrics: Macro-level evaluation compares empirical and generated trajectories using KL divergence, Wasserstein distance, DTW, Macro-F1, and Micro-F1.

D.1. Full Results

MF-MDP achieves the strongest semantic fidelity across the evaluated dimensions by grounding actions in evolving macro state distributions and individual latent states. Its largest advantages appear in state-related semantics and associated sentiment, stance, and behavior dimensions.

  • MF-MDP achieves the largest radar area across long-horizon and reversal settings, indicating the strongest overall semantic fidelity among the compared methods.
  • MF-MDP leads most clearly on State by preserving state-consistent action meaning through explicit coupling of macro sentiment distributions and latent agent states.
  • MF-MDP also improves Sentiment, Stance, and Behavior while maintaining stronger KL, Wasserstein, DTW, Macro-F1, and Micro-F1 alignment.

D.2. Full Analysis

Full dynamic analyses show that MF-MDP tracks long-horizon semantic trajectories and opinion reversals more closely than the baselines. It preserves turning points, post-reversal trends, and balanced state-related behavior across extended simulations.

  • Long-horizon semantic alignment: Across 10,000-step simulations, MF-MDP remains closest to ground truth in both trajectory level and trend across eight semantic dimensions.
  • Long-horizon semantic alignment: MF-MDP preserves positive and negative state trajectories, whereas baselines collapse toward overly neutral or one-sided distributions.
  • State reversals: Across reversal events lasting 6,000 to 40,000 steps, MF-MDP captures turning points and post-reversal trends while remaining close to ground truth.
Loading 2604.05516v2…