acceptodds
Under review as a conference paper at ICLR 2027

PoliticalSim: Multi-Layer Political Dynamics Simulation by LLM Agents

Abstract

Traditional political science analyzed political decision making processes, including voters’ electoral choices, legislative deliberation by selected representatives in parliament, and institutional interaction between parliament and other actors such as the bureaucracy and interest group, by modeling economically rational individuals who act to maximize their utility. These political theories are grounded in mathematically rigorous economic theories such as game theory, neglecting individual's diversity and social interactions, lacking controlled experiments because randomized intervention in real societies is almost impossible. Advances in artificial intelligence have made it possible to reproduce patterns of human interaction that mathematical theory alone could not capture by using Agent-Based Modeling. However, it remains unclear to what extent LLMs can comprehend complex political institutions and replicate human political behavior, and prior work didn't provide a simulation environment for controlled experiments that enable or disable individual mechanisms while holding other conditions constant. In this study, we propose PoliticalSim, a novel framework for simulating the political activities of voters, parliament representatives, and institutional interactions, enabling controlled experiments that enable and disable individual political factors while holding other conditions constant. By this framework, we show that LLMs can reproduce some important phenomena observed in political science in both hypothetical and real world grounded settings, and we experimentally identify which factors affect the final outcomes.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Agent Simulation, LLM Application for Political Science

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sfnsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PoliticalSim, an LLM-based multi-agent framework for simulating political processes, including electoral voting, parliamentary deliberation, and interactions with bureaucracies and interest groups. The framework incorporates mechanisms from political science and allows controlled experiments by enabling or disabling individual factors. The authors evaluate four LLMs using hypothetical elections, real-world election settings across seven countries, and parliamentary simulations. The results demonstrate some consistency with established political science theories and explore how different political mechanisms affect simulated outcomes.

Critical strengths
  1. Comprehensive political simulation framework. The paper integrates electoral voting, parliamentary deliberation, and institutional interactions into a unified LLM-agent framework, extending beyond prior studies that focus on individual political processes.

  2. Mechanism-level analysis. The framework supports controlled ablation experiments to examine the influence of political factors such as party discipline, constituency interests, and interest groups, providing a useful approach for studying simulated political dynamics.

Critical weaknesses
  1. Political behaviors may be directly induced by the prompts rather than emerging naturally. Many expected behaviors are explicitly specified in the agent instructions. For example, voters are encouraged to avoid wasted votes, while legislators are instructed to prioritize constituency interests or follow party discipline. Consequently, the observed agreement with political science theories may primarily reflect instruction following rather than genuine political behavior simulation. The existing prompt ablations do not fully resolve this concern.

  2. Insufficient real-world validation and lack of meaningful baselines. The real-world election results are relatively weak, with the best-performing model correctly identifying the largest party in only two of seven countries. More importantly, the paper lacks comparisons with simple baselines based on historical election results or traditional statistical models, despite providing agents with previous election outcomes and habitual party preferences. It is therefore unclear whether the proposed framework offers meaningful predictive value beyond these inputs.

Questions for authors
  1. How can the authors distinguish political behaviors that genuinely emerge from LLM agents from those directly induced by the prompts? Would the main findings still hold without explicit instructions encouraging strategic voting, constituency prioritization, and party discipline?

  2. Given that agents already receive historical election results and habitual party preferences, how much predictive value does PoliticalSim provide beyond simple historical or statistical baselines? Can the authors provide such comparisons to substantiate the framework's effectiveness?

5 comments

We thank the reviewer for taking time to consider our paper.

First, we clarify that this paper does not claim that political behavior spontaneously emerges from individual LLM agents’ decision making.

The paper addresses two research questions. First, does the ability of LLM agents to emulate human behavior and interactions extend to political simulations? (Abstract, lines 18–22; Introduction, lines 55–58). Second, when simulations are grounded in the real world by directly supplying social media posts or Wikipedia articles, can we distinguish the model’s simulation capabilities from its tendency to reproduce information supplied in the prompt? (Introduction, lines 50–58; Appendix, Figure 11, where “#Ground” denotes methods for real-world grounding).

Regarding the first question, previous research on LLM agent simulations has demonstrated substantial capabilities in emulating human behavior and interactions in domains such as economics [1,2]. However, it is not self-evident that these capabilities extend to the behavior of voters and politicians. One contribution of our study is to investigate this question in the political domain and benchmark LLM models’ ability to simulate political science.

Regarding the second question, simulations of voter opinions that use social media posts as input may produce outputs strongly affected by those input prompts. Such experiments alone do not establish the principle that LLMs can reliably simulate unseen political scenarios for which comparable grounding texts are unavailable. We therefore propose a novel framework that conducts political simulations without detailed textual inputs from social media or Wikipedia. Instead, it grounds the simulations in statistical information as a reference about each country, including religious composition, constituency-level industrial structure, and urban/rural classifications, together with voters’ economic circumstances (Introduction, line 70).

[1] Gati Aher, Rosa I. Arriaga, Adam Tauman Kalai, Using large language models to simulate multiple humans and replicate human subject studies, ICML2023. [2] John J. Horton, Apostolos Filippas, Benjamin S. Manning, Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?, EC '24: Proceedings of the 25th ACM Conference on Economics and Computation.

W1:

Political behaviors may be directly induced by the prompts rather than emerging naturally. Many expected behaviors are explicitly specified in the agent instructions. For example, voters are encouraged to avoid wasted votes, while legislators are instructed to prioritize constituency interests or follow party discipline. Consequently, the observed agreement with political science theories may primarily reflect instruction following rather than genuine political behavior simulation. The existing prompt ablations do not fully resolve this concern.

First, we emphasize that this paper does not claim that political behavior “emerges spontaneously” in LLM agent simulations. Therefore, the criticism concerning emergence does not apply to the claim we make.

Second, the concern that “the observed agreement with political science theories may primarily reflect instruction following rather than genuine political behavior simulation” was already considered in our experimental design and addressed through ablation studies. All conditions specified in the prompts can be individually enabled or disabled, and results with the relevant conditions disabled are reported in Round 1 of Table 1 and in Table 4.

Third, specifying agents’ behavioral principles through prompts and studying the collective outcomes arising from their interactions is common practice in prior work on LLM agent based economic simulations [3,4] and does not compromise the performance or reliability of the simulation framework.

[3] Nian Li, Chen Gao, Mingyu Li, Yong Li, and Qingmin Liao, EconAgent: Large Language Model-Empowered Agents for Simulating Macroeconomic Activities, ACL2024. [4] Yuzhe Yang, Yifei Zhang, Minghao Wu, Kaidi Zhang, Yunmiao Zhang, Honghai Yu, Yan Hu, Benyou Wang, TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets, NeurIPS2025.

W2:

Insufficient real-world validation and lack of meaningful baselines. The real-world election results are relatively weak, with the best-performing model correctly identifying the largest party in only two of seven countries. More importantly, the paper lacks comparisons with simple baselines based on historical election results or traditional statistical models, despite providing agents with previous election outcomes and habitual party preferences. It is therefore unclear whether the proposed framework offers meaningful predictive value beyond these inputs.

Lack of directly comparable baselines:

Previous political simulation frameworks have used approaches such as directly supplying election related social media posts to LLMs [5] or providing detailed candidate profiles constructed from legislators’ Wikipedia articles[6]. These approaches have problems, as the concern raised in Reviewer sfns’s Weakness 1: “Consequently, the observed agreement with political science theories may primarily reflect instruction following rather than genuine political behavior simulation.”

To address this concern, we propose a novel framework grounded in voters’ economic circumstances generated through economic simulations and constituency profiles constructed from national statistics, including religious composition, industrial structure, and urban/rural classifications, that are free from detailed SNS and Wikipedia texts.

To the best of our knowledge, no existing framework provides a directly comparable baseline under the same information constraints. Furthermore, our framework simulates seven countries with substantially different electoral systems, and we are unaware of an existing statistical modeling framework that covers all seven within a comparable experimental setting.

Predictive performance:

Our evaluation considers 4 political science metrics, rather than only whether the model correctly identifies the largest party. Performance should therefore be assessed across all 4 dimensions.

For the best performing model, the rank correlation of party vote totals exceeds 0.66 in more than half of the countries. The simulations also achieve absolute errors within 1.0 on political science measures such as the Gallagher index and the effective number of parties in several national settings.

These results are achieved with limited real-world grounding information and without detailed textual inputs from social media [5] or Wikipedia[6]. We therefore emphasize the framework’s predictive performance under these information constraints.

[5] Maryam Davoodi, Dan Goldwasser, Analysis of State-Level Legislative Process in Enhanced Linguistic and Nationwide Network Contexts, NAACL2024. [6] Hao Li, Ruoyuan Gong, Hao Jiang, Political Actor Agent: Simulating Legislative System for Roll Call Votes Prediction with Large Language Models, AAAI2025.

Q1:

How can the authors distinguish political behaviors that genuinely emerge from LLM agents from those directly induced by the prompts? Would the main findings still hold without explicit instructions encouraging strategic voting, constituency prioritization, and party discipline?

First, we emphasize that we do not claim that political behavior “emerges” from LLM agents.

Second, all experiments include ablations in which prompt conditions are enabled or disabled, allowing us to assess their influence on the simulated outcomes.

Q2:

Given that agents already receive historical election results and habitual party preferences, how much predictive value does PoliticalSim provide beyond simple historical or statistical baselines? Can the authors provide such comparisons to substantiate the framework's effectiveness?

PoliticalSim offers three main contributions: (1) real-world grounding without additional model training or detailed agent profiles; (2) a benchmark for evaluating LLMs’ ability to understand political settings and emulate political behavior; and (3) a platform for investigating political theories while accounting for interactions among agents, particularly where controlled real-world experiments are difficult or infeasible.

(1) Real-world grounding with limited information:

PoliticalSim simulates human political behavior using voters’ economic circumstances generated through economic simulations and limited statistical information, such as religious composition and party support distributions. It does not require additional model training or detailed agent profiles constructed from political related texts written by human such as SNS or Wikipedia. The concern raised in Reviewer sfns’s W1 applies to prior LLM agent simulation approaches for political science that grounding real-world using social media posts [5] or detailed candidate profiles derived from Wikipedia [6]. Our framework reduces reliance on such detailed textual inputs.

(2) Evaluating political understanding and behavioral simulation:

Despite substantial advances in LLM capabilities, their reasoning ability of complex political, legal, and economic settings remains insufficient. PoliticalSim requires models to act not only as voters and legislators but also as bureaucrats and interest groups, enabling evaluation across diverse political roles. For example, Figures 8 and 10 in the Appendix indicate that, when assigned bureaucratic or interest-group roles, Llama-3.1-8B-Instruct makes decisions that are less consistent with maximizing the assigned actor’s interests than those of the other models. Such results illustrate the framework’s potential as a LLM benchmark for assessing political reasoning and role-specific decision-making.

(3) Controlled investigation of political theories:

Comparing competing theories under otherwise identical conditions is often difficult in social science, and some interventions cannot feasibly be conducted in real societies. PoliticalSim provides a platform for examining the implications of theory derived mechanisms under controlled simulated conditions while accounting for interactions among actors. These experiments can support hypothesis development and identify predictions for subsequent empirical testing, although simulation results alone do not establish a theory’s validity in the real world.

[5] Maryam Davoodi, Dan Goldwasser, Analysis of State-Level Legislative Process in Enhanced Linguistic and Nationwide Network Contexts, NAACL2024. [6] Hao Li, Ruoyuan Gong, Hao Jiang, Political Actor Agent: Simulating Legislative System for Roll Call Votes Prediction with Large Language Models, AAAI2025.

Reviewer_wyq3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a new economy-driven political simulation framework called PoliticalSim. It simulates the process of electoral voting, parliamentary deliberation, and institutional interactions among legislatures, bureaucracies, and interest groups, based on mechanisms from political science theories. Overall, this paper has achieved progress in simulating complex political institutions and replicating human political behavior on economical groundings. It examines lower house elections in seven countries which supports the reliability of PoliticalSim. It takes different factors into account and completes ablation study to demonstrate the effectiveness of factors. However, the evidence for real-world validity and discussion on the use of LLMs remains insufficient.

Critical strengths
  1. PoliticalSim integrates electoral voting, parliamentary deliberation, and interactions involving bureaucracies and interest groups in a single framework.
  2. The authors evaluate four LLMs and examine how selected political factors affect simulated outcomes.
  3. There is real-world election experiments that compare simulated outcomes with aggregate election statistics from seven countries.
Critical weaknesses
  1. The paper evaluates four models from different model families. Table 4 reveals divergent results among different models. For example, removing party discipline changes the enactment rate by −11.6 percentage points for Qwen3-14B, but by +18.0 and +20.1 points for GPT-OSS and Gemma. Further explanation should be stated to better illustrate why this happened otherwise the conclusions regarding mechanisms should be weakened. Models behavior analysis among the same model family may help clarify how model-level differences affect the conclusions.
  2. The legislation experiment uses fictional bills. A real-world case study of the reproduction of the legislative amendments and enactment process can help consolidate PoliticalSim.
  3. The paper reports ten runs for the hypothetical election experiments, but it is unclear how many independent runs were conducted for the real-world election and parliamentary experiments. Without the number of runs and variability in key results, it is difficult to assess how robust the findings are across independent runs.
  4. The paper does not fully explain the basis for the weights (0.3 and 0.7) used in Eq. (2). The relationship between the theoretical formula, the prompts, and the observed behavior is also unclear.
  5. The data sources used to determine regional economic disparities in the seven countries should be specified. It would be better if the author could explain how it is ensured that this data aligns with national circumstances and the preferences of local voters.
Questions for authors
  1. Table 4 shows divergent results among models. Could the authors explain why removing party discipline changes the enactment rate by −11.6 percentage points for Qwen3-14B, but by +18.0 and +20.1 points for GPT-OSS and Gemma? Could they also clarify which conclusions about political mechanisms remain robust across models?
  2. Since the legislation experiment uses fictional bills, are there specific real-world cases of legislative amendment or enactment that PoliticalSim can reproduce?
  3. The paper reports ten runs for the hypothetical election experiments. How many independent runs were conducted for the real-world election and parliamentary experiments? Could the authors report the variability of the key results across these runs?
  4. Is there a scientific basis for the weights (0.3 and 0.7) used in Eq. (2) on page 4? Are the main findings robust to other weights? Could the authors clarify the relationship between the theoretical formula, the prompts, and the actual observed behavior?
  5. Could the authors specify the data sources used to determine regional economic disparities in the seven countries and explain how these data align with national circumstances and local voters’ preferences?
4 comments

We thank the reviewer for taking time to consider our paper.

W1&Q1:

The paper evaluates four models from different model families. Table 4 reveals divergent results among different models. For example, removing party discipline changes the enactment rate by −11.6 percentage points for Qwen3-14B, but by +18.0 and +20.1 points for GPT-OSS and Gemma. Further explanation should be stated to better illustrate why this happened otherwise the conclusions regarding mechanisms should be weakened. Models behavior analysis among the same model family may help clarify how model-level differences affect the conclusions.

Our proposed framework elicits explanations for LLM agents’ decisions at each stage, enabling detailed analysis. The observed differences can be explained as follows.

Qwen3-14B agents tend to follow the party line when their party opposes a bill. Consequently, omitting party-line information substantially increases the enactment rate. In contrast, Gemma-4-12b-it and gpt-oss-20b agents consistently consider constituency interests alongside the party line when deciding how to vote.

Table: Proportion of votes in favor of a bill (%) when the legislator’s party opposes it.

ModelParty line oppose & input Party lineParty line oppose & without Party lineParty line support & input Party lineParty line support & without Party line
Qwen3-14B7769992
gpt-oss-20b52869291
Gemma-4-12B17457959
Llama-3.1-8B45899896

For Qwen3-14B, the proportion of votes in favor of a bill is 7% when agents are informed that their party opposes it, compared with 76% when this information is omitted. This substantial difference reflects the model’s general tendency to support bills in the absence of party line information, resulting in a higher enactment rate when that information is removed. The generated explanations explicitly refer to the party line as a consideration in voting decisions.

For Gemma-4-12b-it, support is similarly low (17%) when agents are informed that their party opposes the bill. However, support remains relatively low (45%) when this information is omitted, producing a smaller difference in enactment rates than for Qwen3-14B. The generated explanations suggest that agents consider the interests of their constituencies when party line information is unavailable.

For gpt-oss-20b, support is 52% when agents are informed that their party opposes the bill, compared with 86% when this information is omitted. Thus, even when the party’s opposition is explicitly provided, approximately half of the agents vote in favor of the bill.

When conclusions differ across models, for political LLM agent simulation purposes, we may prioritize models whose behavior more closely matches empirical observations of human behavior. The responses of Gemma-4-12b-it and gpt-oss-20b to the presence and absence of party line information suggest that they may better capture legislators’ consideration of both party positions and constituency interests.

W2&Q2:

The legislation experiment uses fictional bills. A real-world case study of the reproduction of the legislative amendments and enactment process can help consolidate PoliticalSim.

Our research questions are whether LLM agents can emulate human behavior in political settings and, if so, whether they can do so without relying on detailed agent profiles constructed from SNS or Wikipedia texts for real-world grounding. A direct comparison with prior simulation models that improve predictive accuracy through such detailed textual inputs would therefore not constitute a comparison under matched information conditions.

Here we conducted additional simulations using actual 10 bills that passed and 10 bills that failed to pass in each of the United States and France, since whole experiment require 990.7M token LLM reasoning in total. This simulation didn’t use SNS or Wikipedia text as in prior studies, since our aim was to distinguish the effects of overly prescriptive prompts from appropriate real-world grounding. However, our simulation framework achieved 80% accuracy as the highest.

Table: Accuracy (%) of actual bill simulation with real-world use case study.

CountryQwen3-14BGemma-4-12B
USA8065
France8065

W3&Q3:

The paper reports ten runs for the hypothetical election experiments, but it is unclear how many independent runs were conducted for the real-world election and parliamentary experiments. Without the number of runs and variability in key results, it is difficult to assess how robust the findings are across independent runs.

The results of the parliamentary experiments are averaged over seven runs. (line 167)

W4&Q4:

The paper does not fully explain the basis for the weights (0.3 and 0.7) used in Eq. (2). The relationship between the theoretical formula, the prompts, and the observed behavior is also unclear.

The basis for the weights:

The weights in Eq. (2) were chosen for the following reasons.

Drawing on established political science research that identifies reelection as a central motivation for district elected legislators [*1], we assume that legislators place greater weight on constituency interests than on advancement within their party. Political careers are also shaped by incentives to obtain influential party positions [*2]. These considerations motivate our qualitative weighting assumption, rather than uniquely determining the numerical weights. These weights are configurable, allowing researchers to examine how constituency interests and incentives for party advancement affect legislators’ behavior under different assumptions.

The relationship between the theoretical formula, the prompts, and observed behavior:

Eq. (2) calculates each legislator’s individual stake in each bill. These values are supplied in the prompts used for legislative deliberation and institutional interaction stages. The prompts explicitly allow agents to exercise their own judgment. The agents may take these stakes into account, but they are not required to vote in accordance with them. The agents’ stated reasons for their legislative decisions indicate that gemma-4-12b-it and gpt-oss-20b consider maximizing their own interests alongside party discipline.

[*1]

-David R. Mayhew, Congress: The Electoral Connection, Yale University Press, 1974.

-Bruce E. Cain, John A. Ferejohn, and Morris P. Fiorina, The Constituency Service Basis of the Personal Vote for U.S. Representatives and British Members of Parliament, American Political Science Review, 78(1), 110–125, 1984.

-Barry R. Weingast, Kenneth A. Shepsle, and Christopher Johnsen, The Political Economy of Benefits and Costs: A Neoclassical Approach to Distributive Politics,Journal of Political Economy, 89(4), 642–664, 1981.

[*2]

-Giacomo Benedetto and Simon Hix., The Rejected, the Ejected, and the Dejected: Explaining Government Rebels in the 2001–2005 British House of Commons, Comparative Political Studies, 40(7), 755–781, 2007.

-Christopher J. Kam, Party Discipline and Parliamentary Politics, Cambridge University Press, 2009.

W5&Q5:

The data sources used to determine regional economic disparities in the seven countries should be specified. It would be better if the author could explain how it is ensured that this data aligns with national circumstances and the preferences of local voters.

The constituency-level settings for the seven countries were constructed using official statistics published by their respective national agencies, grounding the simulations in documented demographic and socioeconomic conditions. They are all accessible on the internet.

USA: https://www.bls.gov/cew/downloadable-data-files.htm

https://fred.stlouisfed.org/release?rid=140

https://www.ers.usda.gov/data-products/farm-income-and-wealth-statistics/data-files-us-and-state-level-farm-income-and-wealth-statistics

UK:

https://www.ons.gov.uk/file?uri=/economy/grossvalueaddedgva/datasets/nominalandrealregionalgrossvalueaddedbalancedbyindustry/current/regionalgrossvalueaddedbalancedbyindustryandallinternationalterritoriallevelsitlregions.xlsx

https://www.gov.uk/government/statistics/total-income-from-farming-for-the-regions-of-england

eurostat for Germany/France/Italy:

https://ec.europa.eu/eurostat/databrowser/explore/all/all_themes

Cabinet Office in Japan:

https://www.esri.cao.go.jp/jp/sna/data/data_list/kenmin/files/contents/tables/2022/syuyo1.xlsx

Ministry of Agriculture, Forestry and Fisheries in Japan:

https://www.e-stat.go.jp/stat-search/file-download?statInfId=000040156131&fileKind=0

Statistics Korea (National Data Agency):

https://kosis.kr/upsHtml/online.do?isOnline=Y&PART=G&pubcode=GS

https://mods.go.kr/board.es?mid=a10301130300&bid=243&act=view&list_no=442567

Dear Reviewer wyq3,

With the author response deadline approaching, we wanted to ask whether you have any remaining questions or concerns. We’d be happy to clarify any points while the discussion is still open.

If our responses have addressed your concerns, we would appreciate it if you could let us know whether you would consider raising your score.

Thank you for your time and feedback.

Best Regards,

The Authors

AuthorsOpenReview ↗

Gentle Follow-Up on Our Responses

Dear Reviewer wyq3,

Since the author response deadline ends within next 24 hours, we would like to follow up whether you have any remaining questions or concerns. We are still very happy to clarify any points while the discussion is still open.

If our responses have addressed your concerns, we would appreciate it if you could let us know whether you would consider raising your score.

Thank you for your time and feedback.

Best Regards,

The Authors

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

This paper introduces PoliticalSim, an LLM-agent framework that connects voters’ economic circumstances, electoral voting, parliamentary deliberation, and interactions with bureaucracies and interest groups. The framework incorporates mechanisms motivated by political science and allows individual components to be enabled or disabled. The authors evaluate four LLMs in hypothetical elections and settings grounded in seven countries, examining electoral concentration, party representation, legislative outcomes, lobbying, and budget negotiations. They also conduct ablations to investigate how different mechanisms affect simulated outcomes.

Critical strengths
  1. The framework covers a broad range of political processes, bringing together actors and institutional interactions that are often studied separately. This integration is a useful system-building effort.
  2. The modular design supports controlled comparisons within the simulation. Selectively removing mechanisms such as party discipline, constituency interests, and interest-group involvement provides a practical way to examine the framework’s behavior.
  3. The evaluation includes multiple models, electoral systems, and national settings. The detailed prompts, supplementary results, and election-side prompt ablations help readers inspect the modeling assumptions and understand differences across models.
Critical weaknesses

My main concern is that, despite the extensive simulations, the paper does not identify a clear central observation that establishes a sufficiently substantial contribution. Most reported findings qualitatively reproduce familiar political patterns, while some corresponding behaviors are explicitly encouraged by the agents’ role instructions. It therefore remains unclear what new knowledge the simulations provide beyond demonstrating responses to the specified assumptions.

The expanded coverage is valuable, but the experiments do not clearly establish the additional scientific insight enabled by integrating these stages. Similarly, showing that removing a mechanism changes simulation outcomes does not, by itself, demonstrate the claimed ability to discriminate between competing political theories. A more focused research question and a clearer demonstration of the framework’s incremental explanatory value would strengthen the contribution.

Questions for authors

What is the main new finding enabled by the integrated framework, and which experiment most directly establishes it?

6 comments

We thank the reviewer for taking time to consider our paper.

W1:

My main concern is that, despite the extensive simulations, the paper does not identify a clear central observation that establishes a sufficiently substantial contribution. Most reported findings qualitatively reproduce familiar political patterns, while some corresponding behaviors are explicitly encouraged by the agents’ role instructions. It therefore remains unclear what new knowledge the simulations provide beyond demonstrating responses to the specified assumptions.

In general, validating a simulation model against established empirical findings is an important step toward using the simulation model to generate new knowledge and insights [1]. Validation remains a central challenge in LLM based agent simulation. Developing a framework that covers multiple aspects of complex political processes and evaluating its ability to reproduce established findings is therefore a meaningful contribution in its own right.

It remains unclear how reliably LLM based agent simulations can reproduce established political phenomena. LLM’s strong performance on mathematical reasoning tasks does not, by itself, establish that LLMs can understand and reason social science application domains such as political science, legal, and economic fields. LLM ability to reproduce political phenomena through simulation is not self-evident and should therefore be evaluated empirically.

Against this background, our study investigates two research questions and develops a novel framework.

First, does the ability of LLM agents to imitate human behavior and interactions extend to political simulations? (Abstract, lines 18–22; Introduction, lines 55–58). Second, can political simulations achieve real-world grounding without relying on detailed agent profiles constructed from social media posts or Wikipedia articles? Prior researches that achieve real-world grounding by SNS or Wikipedia texts in LLM agent simulation raise the question of whether the resulting behavior reflects the models’ simulation capabilities or primarily reproduces information provided in the prompts (Introduction line 50-58 & Appendix, Figure11, #Ground denotes for real-world grounding methods.).

To investigate these questions, we propose a framework that uses simulated economic circumstances and statistical information to ground agents and evaluate their ability to reproduce established political phenomena.

[1] Windrum, P., Fagiolo, G., & Moneta, A. (2007). Empirical validation of agent-based models: Alternatives and prospects. Journal of Artificial Societies and Social Simulation, 10(2), 8.

W2:

The expanded coverage is valuable, but the experiments do not clearly establish the additional scientific insight enabled by integrating these stages. Similarly, showing that removing a mechanism changes simulation outcomes does not, by itself, demonstrate the claimed ability to discriminate between competing political theories. A more focused research question and a clearer demonstration of the framework’s incremental explanatory value would strengthen the contribution.

Stage Integration:

The interdependence of decisions made by voters, legislators, and bureaucrats is well established in political science [2,3,4,5]. Representing these actors within a unified framework therefore provides a basis for examining how their interactions shape political outcomes.

A narrowly focused research question, as suggested by Reviewer taAx, may overlook an important phenomenon discussed in political science, such as: legislators’ decisions may be shaped less by deliberation with fellow legislators than by factors beyond those discussions, including constituency interests, interest group’s lobbying influence, and prospects for promotion to cabinet positions. Capturing these influences requires a framework that represents interactions across multiple political arenas.

In our simulation, the same constituency information is used to characterize voters’ local contexts, inform legislators’ preferences, and structure interactions between interest groups and legislators. Integrating these stages enables controlled investigations within the simulation of research questions such as whether voters’ preferences affect legislators’ behavior, questions that are difficult to examine through controlled experiments in real-world politics.

Our Claim:

Our study addresses two research questions.

First, does LLMs’ ability to imitate human behavior and interactions extend to political simulations? (Abstract, lines 18–22; Introduction, lines 55–58). Second, can political simulations achieve real-world grounding without relying on detailed textual inputs from social media or Wikipedia? When such texts are supplied directly to LLMs, it can be difficult to distinguish the models’ simulation capabilities from their reproduction of information provided in the prompts (Introduction, lines 50–58; Appendix, Figure 11, where “Ground” denotes real-world grounding methods).

By investigating these questions, our study evaluates the applicability of LLM based agent simulation to political science, which cannot be taken for granted. Furthermore, our simulations of real-world national settings achieve higher accuracy than prior approaches.

Together, these methodological and empirical contributions establish the novelty and significance of our work.

[2] Warren E. Miller, Donald E. Stokes, Constituency Influence in Congress, American Political Science Review, 57(1), pp45–56.

[3] Atif Mian, Amir Sufi, and Francesco Trebbi, The Political Economy of the US Mortgage Default Crisis, American Economic Review, 100(5), pp1967–1998.

[4] Richard L. Hall and Alan V. Deardorff, Lobbying as Legislative Subsidy, American Political Science Review, 100(1), pp69–84.

[5] Kaare Strøm, Delegation and Accountability in Parliamentary Democracies, European Journal of Political Research, 37(3), 261–289.

Q1:

What is the main new finding enabled by the integrated framework, and which experiment most directly establishes it?

A defining feature of social phenomena is that individuals interact with diverse actors across organizational boundaries. Simulating an organization in isolation may therefore overlook interactions with external stakeholders. Legislators’ decisions, for example, are shaped not only by parliamentary debate but also by interest group lobbying, party discipline, cabinet decisions, and negotiations with bureaucrats over policy implementation [4,5,6,7].

Our framework integrates political processes that prior studies have typically examined separately. Constituency profiles incorporating industrial composition and religious demographics provide a shared basis for representing voters’ preferences, constituency elected legislators’ payoff functions, and the behavior of interest groups whose interests align or conflict with those of the constituency. This allows these actors and their interactions to be simulated within a unified framework.

Figure 11 analyzes the reasons that LLM agents report for their legislative voting decisions. Across Qwen3-14B, Llama-3.1-8B-Instruct, Gemma-4-12b-it, and gpt-oss-20b, approximately 10% of agents cite discussions with fellow legislators. By comparison, agents more frequently mention constituency interests relevant to reelection (approximately 65%), local industries (approximately 70%), budget negotiations with bureaucrats or policy implementation costs (approximately 50%), and party discipline (approximately 35%). These reported reasons are consistent with prior political science researches emphasizing influences beyond parliamentary debate.

[4] Richard L. Hall and Alan V. Deardorff, Lobbying as Legislative Subsidy, American Political Science Review, 100(1), pp69–84.

[5] Kaare Strøm, Delegation and Accountability in Parliamentary Democracies, European Journal of Political Research, 37(3), 261–289.

[6] McCubbins, M. D., Noll, R. G., and Weingast, B. R., Administrative Procedures as Instruments of Political Control, Journal of Law, Economics, & Organization, 3(2), 243–277.

[7] Tsebelis G. Decision Making in Political Systems: Veto Players in Presidentialism, Parliamentarism, Multicameralism and Multipartyism. British Journal of Political Science. 1995;25(3):289-325.

Comments on the paper

2 comments

Dear Reviewer,

We hope our response clarify your questions and provide additional insights.

Please feel free to reach out for any further discussion or clarifications!

AuthorsOpenReview ↗

Final Author Remark

We thank the area chair for handling our submission and all reviewers for their feedbacks.

We appreciate that the reviewers recognized our strengths such as broad political process coverage of our framework, the modular design allowing controlled comparisons within the simulation by selectively removing mechanisms, the evaluation including multiple models, electoral systems, and national settings.

We would like to emphasize the following three points:

(1) Regarding Reviewer taAx’s concern that our simulation provides no additional social science insights, we believe that this assessment does not sufficiently account for the contribution of developing and validating an LLM-based simulation framework, as recognized in prior work such as TwinMarket (NeurIPS 2025).

(2) In response to Reviewer wyq3’s concerns, we have demonstrated how our framework supports analysis of agents’ decisions and conducted additional simulations using real-world bills from the United States and France, achieving up to 80% accuracy. These results were obtained without the detailed textual inputs from social media or Wikipedia required by the prior approaches discussed in our response.

(3) Regarding Reviewer sfns’s concern that political behavior does not “naturally emerge,” we clarify that our paper makes no such claim about individual agents’ behavior. Specifying agents’ behavioral principles through prompts and examining the collective outcomes arising from their interactions is an established practice in LLM-based economic simulations (see References 3 and 4 in our response to Reviewer sfns).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.