MF-LLM: Simulating Population Decision Dynamics via a Mean-Field Large Language Model Framework

Dear Reviewer G3Ud,

Thank you very much for taking the time to review MF‑LLM and for your valuable feedback, as well as for recognizing our work with “extensive experiments,” “robust performance improvements,” and “comprehensive analysis providing clear, interpretable evidence of the framework’s effectiveness.”

During the rebuttal, we have made every effort to address your concerns through additional experiments, detailed analysis, and writing refinements:

• Added experiments on diverse datasets — including English Twitter replies and Bitcoin sentiment — showing that MF‑LLM generalizes across languages and diverse forms of collective behavior.
• Provided scalability analysis — formally analyzing complexity reduction from O(N²) to O(N) and demonstrating efficiency gains in large‑scale simulation.
• Extended long‑horizon evaluation — adding experiments up to 2,000 steps to confirm MF‑LLM’s robust performance at scale.
• Refined the contribution summary — presenting the core innovations more concisely to directly address the reviewer’s clarity concerns.

Below, we provide detailed responses to each of your questions, with the utmost sincerity, to address every concern.

---

W1: "The paper claims a general-purpose social simulation framework but is evaluated on only a single Chinese Weibo dataset. As a general method, it should be tested on at least 2–3 diverse datasets. Besides, include at least one English dataset."

Answer: Thank you for this important question. MF-LLM is inherently culture-agnostic, with no assumptions about language or cultural background. To validate its generalization ability, we added experiments on culturally diverse English datasets without domain-specific fine-tuning.

(1) Famous Keyword Twitter Replies Dataset (Kaggle)

This dataset covers globally relevant topics (COVID, Vaccine, NFT, TikTok, WorldCup, etc.). Using Qwen2-1.5B-Instruct, MF-LLM consistently outperforms baselines:

(2) Market Behavior Simulation

To assess MF-LLM’s generalization to market behavior, we refer to TwinMarket [1], which uses real-world transaction data from Chinese social platforms (Xueqiu, Guba), and ElectionSim [2], which uses Twitter data on elections. However, neither releases datasets, making direct comparison infeasible.

Following their approach of using social media data to simulate collective decisions, we use Bitcoin-related tweet–reply data from Twitter to simulate dynamic bullish/bearish sentiment at the population level. Once again, MF-LLM achieves superior alignment:

We hope these additional results address your concern by showing that MF-LLM generalizes across languages and cultural settings. These findings will be incorporated into the revised paper.

[1] TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets.

[2] ElectionSim: Massive Population Election Simulation Powered by LLM Agents.

---

W2: "Scalability is cited as a main motivation, yet the paper lacks any complexity analysis or detailed experiments on efficiency and scalability."

Answer:

(1) Computational Complexity Analysis:

MF-LLM aims to capture dynamic agent interactions, which are fundamental to the formation of collective behavior. Modeling all agent–agent pairwise interactions requires O(N²) complexity, while MF-LLM reduces this to O(N) linear complexity. Specifically, the policy model generates one personalized action per agent, leading to O(N) LLM inferences. The mean field model updates the population signal once per batch of K agents, resulting in O(N/K) mean field inferences.

Overall, MF-LLM reduces total LLM inference complexity to O(N) while still modeling agent interactions and producing population-level dynamics. This contrasts with baselines that either ignore agent interactions entirely or approximate them through manual prompt engineering without dynamic modeling.

(2) Testing on Large Populations:

In the original paper, the Weibo dataset naturally limited simulation horizons to ~900 steps due to the discussion length of real events. To further validate scalability, we extend evaluation to the English Twitter dataset with long-horizon discussions (up to 2,000 steps). Results show that while error increases slightly with horizon, MF-LLM consistently outperforms baselines.

---

W3: "The writing should be improved. For example, the contributions paragraph is overly lengthy and fails to highlight the core innovations succinctly."

Answer: We thank the reviewer for the valuable suggestion. We will further polish the writing in the revised paper. Below is a more concise summary of the contributions, highlighting the core innovations:

(1) MF‑LLM framework — We introduce MF‑LLM, the first framework to integrate mean‑field theory into LLM‑based social simulation, by iteratively updating population signals and individual decisions to capture the evolving trajectories of collective behavior.

(2) IB‑Tune alignment — To improve alignment with real‑world data, we introduce IB‑Tune, an Information Bottleneck–inspired fine‑tuning method that compresses population history into only those population signals most predictive of future actions.

(3) Empirical validation — MF‑LLM achieves up to 47% lower KL divergence than baselines, generalizes across seven domains and four LLM backbones, and supports real‑world forecasting and intervention planning.

---

We hope these clarifications and new results will encourage you to reconsider your score. If you have any further questions or would like additional clarifications, we would be very happy to discuss them.

Dear Reviewer G3Ud,

Once again, we sincerely thank you for the time and effort you have dedicated to reviewing our paper. As the discussion period is approaching its close, we are eager to receive your feedback on our response. We fully understand your busy schedule, but we would greatly appreciate it if you could take our response into account when updating your rating and discussing with the AC and other reviewers.

In addition to the experiments already presented in the rebuttal, including:

• Experiments on English datasets (Twitter) and market behavior simulation,
• Simulation experiments extending to 2000 steps,
• Computational Complexity Analysis,

we have added further experiments to strengthen our claims:

---

W1: "The paper claims a general-purpose social simulation framework but is evaluated on only a single Chinese Weibo dataset. As a general method, it should be tested on at least 2–3 diverse datasets. Besides, include at least one English dataset."

We conducted additional evaluations of MF‑LLM (without fine‑tuning) on the Twitter dataset using GPT‑4o‑mini, GPT‑4o, and DeepSeek‑R1‑Distill‑Qwen‑32B backbones. Results show that MF‑LLM consistently generalizes across cultural contexts.

---

W2: "Scalability is cited as a main motivation, yet the paper lacks any complexity analysis or detailed experiments on efficiency and scalability."

(1) Detailed experiments on Scalability

We also extended the simulation length (Qwen2‑1.5B‑Instruct, without fine‑tuning) to the equivalent of 4000 agent updates (previously reported as “steps”). Results for 500–4000 agents are shown below, confirming stable performance over long horizons.

(2) Detailed experiments on complexity analysis

To complement the theoretical O(N) complexity analysis presented earlier, we conducted empirical measurements to verify runtime scaling in practice. We measured simulation runtime for varying numbers of agent updates on Qwen2‑1.5B‑Instruct (Twitter dataset, without fine‑tuning). Experiments were run on a single NVIDIA A100 GPU (40 GB memory).

Finding: Runtime increases linearly with the number of agents, maintaining a nearly constant time per agent (~0.56 s), which primarily reflects the cost of repeated LLM calls. This empirical result validates the theoretical O(N) complexity, providing experimental evidence for MF‑LLM’s scalability.

---

We hope that these efforts will alleviate your concerns regarding MF‑LLM. Your feedback is highly valuable to us, and we would appreciate any updates or further guidance you might have regarding our revisions and responses.

Dear Reviewer U3Se,

Thank you very much for taking the time to review MF‑LLM and for your constructive feedback, as well as for recognizing the “interesting idea” and “better fit.”

During the rebuttal period, we have made every effort to address your concerns through targeted clarifications, additional experiments, and concrete examples:

• Clarified the simulation mechanism and connectivity.
• Added new case studies on Different Interactions and Different Users.
• Provided explicit mean field summaries, showing agent actions and corresponding updated mean field content.
• Demonstrated that the states in Eq. (1) are observable from real‑world data, not unobservable “internal states.”
• Conducted control experiments by adding state information to baselines, with MF‑LLM still outperforming all variants.
• Reported statistical significance with mean ± standard deviation over multiple runs.

Below, we provide detailed responses to each of your questions, with the utmost sincerity, to fully address every concern.

---

W1 & Q2: Is the mean field summary used to replace agent-agent interactions or in addition to the interactions listed under 'baselines'? Could you provide examples of mean field summaries?

Answer: We thank the reviewer for the insightful comments. We will clarify the simulation description in the revised paper.

1. Mechanism:

Many existing works model the connectivity of the agent network, where each edge weight reflects the strength of influence between two agents. While such network modeling is reasonable, manually specifying these edge weights is often inaccurate and may overlook important interaction patterns in unseen scenarios.

MF‑LLM replaces fixed, handcrafted connectivity with a data-driven mechanism:

• The policy model generates individual actions based on the agent’s state and the current population signal.
• The mean field model updates the population signal using these newly generated actions

This creates an indirect but generalizable interaction pathway: $a_A→m_{t+1}→a_B$ where agent A’s action affects agent B via the updated population signal. Crucially, interaction strengths are inferred and updated dynamically at each step from online data, without any manually set network weights.

The baselines represent alternative agent interaction strategies:

• Recent connects agents to the most recent actions.
• Popular connects agents to the most popular actions.

These fixed heuristics may miss subtle interactions in new contexts. In contrast, MF-LLM allows interaction patterns to emerge adaptively from the data.

2. Experimental Evidence:

Our MF-LLM do not force the mean field model to attend to or ignore specific agents. It evaluates all available information to infer interaction effects. The following two case studies illustrate this dynamic modeling, while also showcasing the content of the mean field.

(1) Influence of Different Interactions

In a rumor propagation scenario, we compare the effects of two different actions taken by the same high-influence user.

• State: A user with high influence and many friends.

• Previous Mean Field:

  Most users express opposition and criticism toward the event, regarding it as unfair competition.

• Action A: "Repost"

• Action B: "Looking forward to the truth!"

The mean field model takes the individual’s state and action as input and updates the mean field accordingly:

Updated Mean field:

• After Action A:

  "Most users are criticizing the event rather than seeking factual verification."

• After Action B:

  "Users call for an investigation to uncover the cause. … Many suspect that this might be a genuine news report."

Due to rebuttal length limits, only excerpts of updated mean field are shown here; the full version will appear in the revised paper.

(2) Influence of Different Users

In the rumor propagation scenario, we compare the effects of the same action taken by two users: one with high influence and one with low influence.

• Action: "Looking forward to the truth!"

• Previous Mean Field: (Same as in Case Study 1)

• Agent States:

  • High-Influence User: A user with high influence and many friends.
  • Low-Influence User: Same description, but with low influence and few friends.

Updated Mean Field:

• After High-Influence User acts: "Many suspect that this might be a genuine news report."

• After Low-Influence User acts: "A few users in the comments express a desire to uncover the truth."

Conclusion: MF‑LLM thus provides an indirect yet generalizable mechanism for modeling agent interactions and influence, without manually specifying network weights.

---

W2 & Q1: "How do we know whether the reported improvements in predictive fit might be because of this new and unrealistic form of interaction?"

Answer: We thank the reviewer for raising this concern.

1. Observed states in Equation 1. The “state” in Eq. (1) corresponds to information observable in the real-world social interactions being modeled. These states include attributes such as location, description, gender, friend/follower counts, activity level, and verification status. The User State Construction prompt template is shown in Appendix D.4 (p. 20) of the submitted paper. No hidden or unobservable “internal states” are introduced.

2. The mean field model processes these observable states, which is fully reasonable. As illustrated in above Case Study 2 (Influence of Different Users), even when different users perform the same action, their states lead to different downstream effects on subsequent users. Processing these observable states enables the mean field model to analyze the corresponding influence of actions more accurately.

3. We conducted a control experiment by adding user state information to the baselines (Recent, Popular). Recent (+state) improves, while Popular (+state) drops in performance. MF-LLM still outperforms these enhanced baselines, confirming that the improvement comes from its interaction modeling. Baselines ignoring user identity fail to capture heterogeneous agent influence, highlighting MF-LLM’s advantage.

---

Q3: "Is there any estimate of the uncertainty and statistical significance of the results in tables 1,2?"

Answer: We report mean ± standard deviation and p-values for Tables 1 and 2. The small standard deviations and consistent performance of MF‑LLM indicate statistical significance and stability.

Due to rebuttal length limits, we only present the main results here; the full version will be included in the revised version.

Tables 1 - mean ± std:

Due to character limits, we present the KL Divergence statistical significance for the Qwen2-1.5B-Instruct model. The following p-values highlight significant differences between algorithms:

(*) p-values < 0.05 indicate statistically significant differences.

Tables 2:

---

We hope these clarifications and new results will encourage you to reconsider your score. If you have any further questions, we are happy to discuss them.

Dear Reviewer GZm3,

Thank you for your detailed and constructive review, and for recognizing our work with a “strong and well‑structured” technical design, “complete and convincing” experiments, and a “clear and well‑organized” paper.

During the rebuttal, we have made every effort to address your comments through additional experiments, clarifications, and expanded analysis:

• Added case studies of highly influential individuals (Appendix C, Fig. 8‑4) and small, close‑knit groups, demonstrating that MF‑LLM effectively captures complex interactions without manual design.
• Added new decoding experiments (varying temperature and top‑p) and highlighted Appendix G outputs showing lexical homogeneity in larger LLMs.
• Conducted sensitivity analysis over multiple β values, showing the relationship between β and loss values.
• Demonstrated that longer warm‑up phases improve long‑term simulation fidelity (Appendix C, Fig. 5a).

Below, we provide detailed responses to each of your questions, with the utmost sincerity, to address every concern.

---

W1 & Q1: "Mean-Field Approximation ... may miss more complex and local interactions"?

Answer: We respectfully clarify that MF‑LLM does not “miss more complex and local interactions” or “break down” in such cases. MF‑LLM is evaluated on real‑world data, which naturally includes "small, close‑knit groups" and "highly influential individuals". We address this via mechanism and empirical evidence:

1. Mechanism

In MF‑LLM, the policy model maps population signals to individual actions, while the mean field model updates the population signal from new actions. The pathway $a_A→m_{t+1}→a_B$ represents an indirect but generalizable interaction between any pair of agents. Unlike prior work with fixed influence weights, MF‑LLM infers and updates interaction effects dynamically from online data, enabling robust modeling even in unseen structures.

2. Empirical Evidence

(a) Highly influential individuals
Appendix C and Fig. 8(4) show "impact of highly influential individuals": when the famous Chinese sports commentator "@Huang Jianxiang" was referenced to verify the truth, the population signal shifted toward skepticism, leading to reduced rumor spread.

(b) Small, close‑knit groups
In an entertainment discussion:

• User A: "Taylor Swift will hold a concert!"
• User B: "The city marathon route will be adjusted next month."

Updated Mean Field (after A+B):
"Population sentiment surges around the concert... marathon news draws moderate interest from runners."

Policy model responses under same mean field:

• Idol Fan User: "Can’t believe it! Already organizing a group to buy tickets and cheer!"
• Marathon Enthusiast: "Good to know about the route change—I’ll adjust my training."

This shows MF‑LLM captures topic‑specific shifts via the mean field and generates heterogeneous actions via the policy model—faithfully modeling "close‑knit group" dynamics from real data.

---

W2: Although the authors show that adding external signals improves simulation accuracy, how does the model handle the automatic detection in real-world scenarios?

Answer: We appreciate the reviewer’s recognition of our “honest and clear” limitation discussion. Appendix C & Fig. 8 shows that adding external signals preserves simulation fidelity, demonstrating MF‑LLM’s robustness. While automatic detection is beyond our current simulation-focused scope, we agree it is important for real-world forecasting and identify it as a future direction.

---

W3 & Q2: Can you provide more evidence for why smaller LLMs outperform larger ones, such as by exploring decoding parameters?

Answer: Regarding this "interesting and unexpected result", we discussed it in Appendix G and presented a comparison of model outputs (pp. 27–28 for larger LLMs vs. pp. 25–26 for smaller LLMs). The larger LLMs’ generations exhibit clear lexical homogeneity—frequent repetition of similar phrasing and syntactic structures (highlighted by color)—whereas the smaller LLMs, under the same simulation context, show no such degree of lexical uniformity.

Further, we conducted additional experiments varying decoding parameters for the larger models. The following table presents the results for:

1. Higher temperature (temp = 1.2, top‑p = 0.85)
2. Higher top‑p (temp = 0.8, top‑p = 0.95)

For each metric, we report the absolute value together with its relative change (%) compared to the original decoding setup in the submitted paper (temperature = 0.8, top‑p = 0.85).

7B = Qwen2‑7B‑Instruct

These results show that increasing diversity through higher temperature or top‑p does not necessarily improve simulation performance. For example, Qwen2‑7B‑Instruct showed only a marginal improvement when increasing temperature, but significant degradation at top‑p = 0.95 (KL more than doubled). Similarly, GPT‑4o‑mini exhibited worse performance when either temperature or top‑p was increased.

Due to rebuttal time constraints, we report the additional results above. More extensive experiments on “increasing diversity” will be included in the revised paper for reference.

---

Q3: How sensitive is the model's performance to the β in the IB-Tune objective?

Answer: We thank the reviewer for raising the question regarding the sensitivity of hyperparameter β in balancing compression and prediction.

In our experiments, we tested β values of 1.2, 1.6, and 2.0 and observed that the predictive term $- \sum_{i=1}^{N_t} \log \pi(a_t^{i} \mid s_t^i, m_t)$ (denoted as $L_2$ hereafter) in Equation 7 shows a qualitative negative correlation with β. The following table summarizes the average $L_2$ over the final 100 steps of training:

As shown, a lower $L_2$ indicates better predictive performance. Based on these qualitative observations, we selected β = 2 for the final model.

We have already begun testing larger β values in additional experiments. However, due to the extensive fine-tuning time required for LLMs, we aim to obtain results before the discussion period ends. We will include these findings in the revised version of the paper.

---

Q4: How does the warm-up duration affect the long-term fidelity of the simulation?

Answer: We thank the reviewer for the insightful question regarding the warm‑up phase.

As discussed in Appendix C and shown in Fig. 5(a), we evaluated simulations starting at different warm‑up lengths (20, 34, 43, 70 steps). Longer warm‑up phases consistently yielded smaller deviation from the real‑world trajectory. The figure also marked the prediction and real values, directly illustrating this improvement. This benefit arose because a longer warm‑up provided the simulation with a richer and more accurate prior, reducing early‑stage uncertainty.

---

We hope these clarifications and new results will address your concerns. If you have any further questions, we are happy to discuss them.

Dear Reviewer f92d,

Thank you for your detailed review and valuable feedback, and for recognizing MF‑LLM as a “theoretically grounded” approach with “robust performance.”

During the rebuttal, we conducted extensive experiments and analyses to address your comments:

• Added experiments on culturally diverse English datasets (Twitter) and market behavior simulation, confirming MF-LLM’s generalization cross different cultural contexts.
• Provided mechanism explanation, empirical results (Appendix C, Fig.8) and new case studies, showing MF-LLM handles heterogeneous interactions and disproportionate influence.
• Added comparison of GPT‑4o‑mini with human experts, GPT-4, GPT-4o, GPT-3.5, confirming it is the best balance of accuracy, cost, and speed for semantic evaluation.
• Added computational complexity analysis (O(N²)→O(N)) and long-horizon experiments (Twitter, 2000 steps), demonstrating scalability.

Here are our detailed responses to each of your questions and suggestions. We have made every effort, with the utmost sincerity, to address every one of your concerns.

---

Weakness & Q2: "How would you validate MF-LLM's performance on entirely different cultural contexts where social interaction norms differ significantly?"

Answer: Thank you for this important question. MF-LLM is inherently culture-agnostic, with no assumptions about language or cultural background. To validate its generalization ability, we added experiments on culturally diverse English datasets without domain-specific fine-tuning.

1. Famous Keyword Twitter Replies Dataset (Kaggle)

This dataset covers globally relevant topics (COVID, Vaccine, NFT, TikTok, WorldCup, etc.). Using Qwen2-1.5B-Instruct, MF-LLM consistently outperforms baselines:

2. Market Behavior Simulation

To assess MF-LLM’s generalization to market behavior, we refer to TwinMarket [1], which uses real-world transaction data from Chinese social platforms (Xueqiu, Guba), and ElectionSim [2], which uses Twitter data on elections. However, neither releases datasets, making direct comparison infeasible.

Following their approach of using social media data to simulate collective decisions, we use Bitcoin-related tweet–reply data from Twitter to simulate dynamic bullish/bearish sentiment at the population level. Once again, MF-LLM achieves superior alignment:

We hope these additional results address your concern by showing that MF-LLM generalizes across languages and cultural settings. These findings will be incorporated into the revised paper.

[1] TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets.

[2] ElectionSim: Massive Population Election Simulation Powered by LLM Agents.

---

Q1: "How does the mean field approximation handle scenarios where agent interactions are highly heterogeneous or when certain agents have disproportionate influence on population dynamics?"

Answer: First, we clarify that MF-LLM is evaluated on real-world data, which naturally involves abundant heterogeneous agent interactions and cases of disproportionate influence. Below, we address this from both a mechanism and empirical evidence perspective:

1. Mechanism: MF-LLM places no assumptions on agent homogeneity. At each step, the policy model generates personalized actions, while the mean field model (an LLM) infers how these actions affect collective dynamics and outputs the key population signals (=mean field) that influences future individual decisions. Unlike prior methods that manually encode interaction heterogeneity or assign fixed influence weights, our approach automatically generates these effects at each step.

2. Empirical: Experimental results in Appendix C and Fig. 8 (4) have already demonstrated disproportionate influence: when the famous Chinese sports commentator "@Huang Jianxiang" was referenced to verify the truth, the population signal shifted toward skepticism, leading to reduced rumor spread.

3. We added two sets of controlled experiments to demonstrate that the mean field model can capture heterogeneous interactions and disproportionate influence without any manual weighting or structural assumptions:

(1) Heterogeneous Interactions

We compare the effects of two different actions taken by the same high-influence user.

• State: A non-verified user with high influence and many friends.

• Previous Mean Field:

  Most users express opposition and criticism toward the event, regarding it as unfair competition.

• Action A: "Repost"

• Action B: "Looking forward to the truth!"

The mean field model takes the individual’s state and action as input and updates the mean field accordingly:

• After Action A:

  "Most users are criticizing the event rather than seeking factual verification."

• After Action B:

  "Users call for an investigation to uncover the cause. … Many suspect that this might be a genuine news report."

Due to rebuttal length limits, only excerpts of updated mean field are shown here; the full version will appear in the revised paper.

(2) Disproportionate Influence

We compare the effects of the same action taken by two users: one with high influence and one with low influence.

• Action: "Looking forward to the truth!"

• Previous Mean Field: (Same as in Case Study 1)

• Agent States:

• • High-Influence User: A non-verified user with high influence and many friends.
  • Low-Influence User: Same description, but with low influence and few friends.

Updated Mean Field:

• After High-Influence User acts: "Many suspect that this might be a genuine news report."

• After Low-Influence User acts: "A few users in the comments express a desire to uncover the truth."

---

Q3: "How do you address potential bias from using GPT-4o-mini for semantic evaluation? ..."

Answer: To balance evaluation cost and speed, we chose GPT-4o-mini for semantic evaluation. We validated its reliability through human evaluation on 100 samples by domain experts (social science scholars, junior faculty, and senior PhD students), comparing GPT-4o-mini’s outputs with human judgments. We also tested other LLMs (GPT-3.5-Turbo, GPT-4o, GPT-4). Summary:

Since semantic evaluation is not a challenging task for today’s advanced LLMs, GPT-4o-mini balances accuracy and cost effectively, offering higher human alignment than GPT-3.5-Turbo at a much lower cost. With an optimized prompt (Appendix D.3), GPT-4o-mini provides the best cost-effectiveness, especially considering its low price, making it ideal for large-scale simulations. We will include full results in the revised paper.

---

Limitation: "computational complexity analysis and testing on truly large populations."

Answer:

1. Computational Complexity Analysis:

MF-LLM aims to capture dynamic agent interactions, which are fundamental to the formation of collective behavior. Modeling all agent–agent pairwise interactions requires O(N²) complexity, while MF-LLM reduces this to O(N) linear complexity. Specifically, the policy model generates one personalized action per agent, leading to O(N) LLM inferences. The mean field model updates the population signal once per batch of K agents, resulting in O(N/K) mean field inferences.

Overall, MF-LLM reduces total LLM inference complexity to O(N) while still modeling agent interactions and producing population-level dynamics. This contrasts with baselines that either ignore agent interactions entirely or approximate them through manual prompt engineering without dynamic modeling.

2. Testing on Large Populations:

In the original paper, the Weibo dataset naturally limited simulation horizons to ~900 steps due to the discussion length of real events. To further validate scalability, we extend evaluation to the English Twitter dataset with long-horizon discussions (up to 2,000 steps). Results show that while error increases slightly with horizon, MF-LLM consistently outperforms baselines.

---

We hope these clarifications and new results will encourage you to reconsider your score. If you have any further questions, we are happy to discuss them.

Dear Reviewer f92d,

Once again, we sincerely thank you for the time and effort you have dedicated to reviewing our paper. As the discussion period is approaching its close, we are eager to receive your feedback on our response. We fully understand your busy schedule, but we would greatly appreciate it if you could take our response into account when updating your rating and discussing with the AC and other reviewers.

In addition to the experiments already presented in the rebuttal, including:

• Experiments on English datasets (Twitter) and market behavior simulation,
• Case studies of heterogeneous interactions and disproportionate influence,
• Comparison of GPT‑4o‑mini for semantic evaluation with human experts, GPT‑4, GPT‑4o, and GPT‑3.5,
• Simulation experiments extending to 2000 steps,

we have added further experiments to strengthen our claims:

---

Weakness & Q2: "How would you validate MF-LLM's performance on entirely different cultural contexts where social interaction norms differ significantly?"

We conducted additional evaluations of MF‑LLM (without fine‑tuning) on the Twitter dataset using GPT‑4o‑mini, GPT‑4o, and DeepSeek‑R1‑Distill‑Qwen‑32B backbones. Results show that MF‑LLM consistently generalizes across cultural contexts.

---

Limitation: "computational complexity analysis and testing on truly large populations."

(1) Testing on Large Populations

We also extended the simulation length (Qwen2‑1.5B‑Instruct, without fine‑tuning) to the equivalent of 4000 agent updates (previously reported as “steps”). Results for 500–4000 agents are shown below, confirming stable performance over long horizons.

(2) Runtime complexity

To complement the theoretical O(N) complexity analysis presented earlier, we conducted empirical measurements to verify runtime scaling in practice. We measured simulation runtime for varying numbers of agent updates on Qwen2‑1.5B‑Instruct (Twitter dataset, without fine‑tuning). Experiments were run on a single NVIDIA A100 GPU (40 GB memory).

Finding: Runtime increases linearly with the number of agents, maintaining a nearly constant time per agent (~0.56 s), which primarily reflects the cost of repeated LLM calls. This empirical result validates the theoretical O(N) complexity, providing experimental evidence for MF‑LLM’s scalability.

---

We hope that these efforts will alleviate your concerns regarding MF‑LLM. Your feedback is highly valuable to us, and we would appreciate any updates or further guidance you might have regarding our revisions and responses.

Dear reviewer f92d,

As the reviewer–author discussion deadline is approaching, we will be online waiting for your feedback on our rebuttal, which we believe has fully addressed your concerns.

We would highly appreciate it if you could take into account our response when updating the rating and having discussions with AC and other reviewers.

Thank you so much for your time and efforts. Sorry for our repetitive messages, but we're eager to ensure everything is addressed.

Authors of Paper #4199

Dear Reviewer f92d,

We sincerely thank you for your valuable comments on improving our work. We would like to respectfully highlight the updates made in response to your suggestions and to address your concerns:

1. Added evaluations on diverse datasets (English Twitter, market simulation) using Qwen2-1.5B-Instruct, GPT-4o-mini, GPT-4o, and DeepSeek-R1-32B, validating MF-LLM’s performance across entirely different cultural contexts.
2. Clarified the mechanism with mean-field approximation,and provided empirical evidence (Appendix C, Fig. 8) and two sets of controlled experiments, showing MF-LLM’s ability to handle heterogeneous interactions and disproportionate influence.
3. Added comparison of GPT-4o-mini with human experts, GPT-4, GPT-4o, and GPT-3.5, confirming it offers the best trade-off between accuracy, cost, and speed for semantic evaluation.
4. Added computational complexity analysis (O(N²) → O(N)) and empirical runtime measurements, verifying linear scaling in practice.
5. Extended to larger populations (up to 4000 agents) and showed that the error increases only marginally from 500 to 4000 agents.

We respectfully hope that you will take these updates and our responses into account when making the final decision. As the author–reviewer discussion period is coming to an end, we look forward to further engaging with you. We would be happy to continue the discussion and provide additional information if you have any further concerns or questions.

Authors of Paper #4199