acceptodds
Under review as a conference paper at ICLR 2027

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

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 **M²Sim**, a social simulation framework that tightly couples **m**acro-level collective dynamics with **m**icro-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 opinion states 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, M²Sim supports stable simulation of long-horizon social events with up to 40,000 interactions (compared to  300 in the baseline MF-LLM), 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.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Social Network Simulation, Large Language Models, Mean-Field Modeling, Opinion Dynamics, Long-Horizon Simulation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_mucgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes M²Sim, a macro-micro coupled framework for long-horizon social simulation with LLM-based agents. M²Sim introduces explicit macro-level state distributions and micro-level latent agent states, enabling the simulation of both population dynamics and individual-level evolution. Experiments demonstrate that M²Sim improves over existing baselines in long-horizon simulation and opinion reversal prediction.

Critical strengths
  1. The paper addresses an important and challenging problem in LLM-based social simulation. The authors identify the limitation of existing mean-field LLM simulators in capturing long-term opinion dynamics and propose a macro-micro coupled framework to model both population-level evolution and individual-level states.

  2. The proposed framework is well motivated. The theoretical analysis provides valuable insights into the failure modes of MF-LLM. By characterizing the self-strengthening behavior of majority states, the paper offers a principled motivation for introducing explicit micro-level states and macro-level dynamics. The experimental results on long-horizon and opinion reversal scenarios demonstrate promising improvements over existing baselines.

Critical weaknesses

Major Weaknesses

  1. The experimental comparison is insufficient. The current experiments mainly compare against Direct LLM, Social Retrieval, and MF-LLM, which are relatively outdated baselines. For an ICLR-level paper, the authors should include more recent LLM-based social simulators to better demonstrate the advantages of M2SimM^2Sim over existing approaches.

  2. The ablation study is too coarse to fully understand the contribution of each component. Finer-grained analysis is needed for key design choices, such as the necessity and effectiveness of the K-step rollout consistency loss.

  3. The Policy Model Optimization relies on the state transition model for future trajectory evaluation, but the counterfactual capability of this transition model is not sufficiently validated. M2SimM^2Sim assumes that different agent actions can be reliably evaluated through the learned transition model fψf_\psi to predict future macro dynamics. However, the experiments do not analyze whether the transition model can accurately capture action-conditioned dynamics or whether policy optimization is affected by potential transition model bias.

  4. The Introduction lacks clarity and accessibility. It introduces many domain-specific terms from social simulation, such as variance damping, majority lock-in, and delayed evidence accumulation, which make the motivation difficult to follow even for readers familiar with the field. Moreover, the connection between the limitations of MF-LLM and the design choices of M2SimM^2Sim requires a clearer clarification.

Minor Weaknesses

  1. The placement of Figure 1 may affect readability. The current framework figure contains many components, including macro dynamics, micro states, and policy optimization. However, readers may not yet understand these concepts when encountering the figure in the Introduction. The authors may consider moving it to the Method section or simplifying the figure.

  2. The main paper lacks a dedicated Related Work section.

  3. Minor formatting issues: Section 2.3 has an inconsistent title format, with an unnecessary period at the end of the section title. In addition, “MF-MDP” in Figure 2 appears to be the name used in an earlier version of the paper.

Questions for authors

See weaknesses

Reviewer_nTJFOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents M²Sim, an LLM-based framework for long-horizon social network simulation. It couples micro-level agent opinion states with a macro-level distribution of collective opinions, using a learned state-transition model and multi-step rollout-based action selection to improve trajectory stability and capture opinion reversals. Experiments on Weibo data and additional reversal events show improved alignment with real-world trajectories over several baselines, particularly in long-horizon and reversal settings.

Critical strengths
  1. The paper adopts an intuitive training approach that fits the model to long-horizon prediction tasks. This provides a way to reproduce real-world dynamics in simulation and has practical value.

  2. The experiments cover several settings: short-horizon, long-horizon, and opinion-reversal simulations. The paper compares M²Sim with multiple LLM-based social simulation baselines and reports notably better trajectory metrics in the long-horizon and reversal settings.

  3. The ablation study focuses on the training components. By removing state-transition training, policy training, and candidate sampling in turn, the authors analyze each component’s effect and provide evidence for their contributions.

Critical weaknesses

The paper has serious problems in how it presents its contributions. The first contribution begins by claiming “a new social dynamic simulation framework,” yet the framework builds on MF-LLM’s existing loop: population-level information, individual actions, updated population-level information, the next round of individual actions. I therefore do not consider the first contribution valid as stated.

The second contribution is also puzzling. Many prior papers already model individual states. For example, TrendSim explicitly maintains emotions, opinions, and social confidence, updating them through reflection after interactions. HiSim’s ordinary users maintain continuous attitude values that are updated based on incoming information. These works already incorporate state transitions, yet neither is cited or analyzed. Opinion reversal is also a problem directly studied by FDE-LLM. Overall, the contribution section fails to identify the paper’s actual contribution.

If state transitions are intended to be the central contribution, the authors should compare against existing methods that combine LLMs with explicit dynamical models as baselines. They should also report the computational costs of the experiments.

Questions for authors

Why you put related work in Appendix? I can't separate your contribution, and your claim is misleading to the readers.

Reviewer_iPAnOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces a new framework for social network simulation that couples macro-scale dynamics with agent states. The authors introduce an explicit distribution of opinion states alongside a textual summary of collective behavior. The framework consists of two main components: (1) a Transformer-based state transition model that predicts the evolution of the macro-level opinion distribution, and (2) an policy model that generates individual actions. Both components are optimized through Long-horizon Consistency Training (LCT). The method is evaluated on real-world datasets and the authors report improvements over existing baselines, particularly in long-horizon simulations.

Critical strengths
  • S1. The research direction is relevant, particularly the difficulty of simulating collective behavior over long periods.
  • S2. The idea of explicitly modeling opinion distributions alongside textual summaries seems to improve the accuracy of simulations.
  • S3. The reported experimental results are promising, especially for long-horizon simulations, where the method outperforms the baselines.
Critical weaknesses

Major

  • W1. The theoretical motivation does not establish the claimed self-strengthening behavior. In Remark 2.2, the authors claim a majority lock-in phenomenon. However, Proposition 2.1 only claims that, given its assumptions, the current majority (at round t) state's fraction does not decrease in expectation. It does not guarantee either that its fraction increases in a particular realization or that it remains the majority. This means that, across three rounds, the current majority could cease to be the majority. Moreover, it is not clear how it is linked to previous methods (e.g. MF-LLM).
  • W2. The state transition model is underspecified. The authors introduce a training method combining two losses in order to train the state transition model f_\psi. However, in the rollout consistency objective, it is not specified (1) how the authors compute r_{t+k-1}, p_{t+k}, c_{t+k}: are they part of the dataset? if so, please specify it. If not, how are they computed? (2) f_\psi does not output the state of each single agent (row 248), yet the action-policy defined in Eq. 5 uses the state of the agent to generate the action, which is then used to generate the textual synopsis r_t. Moreover, it is not clear how the r_t summarizer is implemented.
  • W3. The policy optimization procedure is not adequately defined. Equation (12) introduces a predicted future macro distribution that depends on the candidate actions, policy, and dropout configuration. However, this dependency is not formally defined. In particular, it is unclear how candidate actions affect future macro distributions.
  • W4. The authors did not release the code.

Minor

  • W1. The background section lacks citations. In Section 2.2, you introduce a "meta-dynamic" that the previous methods use. However, these methods are not cited in this section (The background section 2 does not have citations at all).
  • W2. The main difference between the proposed methods and the previous ones is that yours has a mean-field attribute m_t that tracks the macro state. This should be stated clearly in the background section.
  • W3. Inconsistent notation. Notation is inconsistent across the whole paper. Often, after a concept's definition, the authors change it's meaning without explicitly mentioning it. To exemplify, the definition of mean-field summarizer given at line 139 is not the same as the definition given at row 150 (eq.1). Another example is f_\psi, which takes the states x_1,...,x_t as input in eq.7 and only x_t in eq. 8.
  • W4. The authors often mention "active agents" throughout the manuscript. However, this term is never defined.
  • W5. Proposition 2.1 has no proof in the main manuscript. Please provide a link to the appendix.
Questions for authors
  1. How does the model f_\psi generate agent states?
  2. In Equation (8), how are the future inputs obtained?
  3. In Equation (12), how is (\hat m_{t+k}(\pi_\phi(\vec a),\lambda)) computed?

Comments on the paper

1 comment
~Rong_huan1OpenReview ↗

Learning Mechanism is with Originality

Indeed, this paer has referred MF overall scheme, yet there still exists some novelty in terms of the learning mechanism, which should be acknowledged.

  1. MF only return mean-field dynamics signal before each round of content generation. However, for M^2Sim, the dynamic signal is oriented to the long-term rolling out. That is, we predict the most possible next state from the current mean-field statuts, so as to better process long-term simulation.

  2. MF has not considered individual preference on cognition. In this paper, each agent has been assigend with an individual signal to retain diversity, along with the global macro-signal to target the overall evolving direction. Such the dual-enhacenment is with the most value of this paper compared with the current works.

  3. From prompt to representation. Keeping the transformer model-body, this work maintains the learning on dyanmic-signals at the representation level, which can incorporate more interpretability compared with the prompt-text alone.

In my view, this paper should be further polished: ① modeling on field dynamics. what is the prototype in mathematics ② signal evolving coupled with the multi-agent field evovling. How the guide the evovling direction for agents ③ check on individual diversity. If the injection of macro-signal will erase the individual preference to be over homogeneous.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.