FlockVote: LLM-Driven Election Simulation

• FlockVote is an LLM-empowered agent-based modeling framework that simulates electoral behavior by assigning high-fidelity demographic personas to synthetic voters.
• The framework integrates empirical demographic data with contextual election issues to prompt agents into generating probabilistic voting decisions, balancing interpretability with predictive replication.
• FlockVote offers an experimental laboratory for analyzing agent-level rationales and sensitivity to prompts, bridging traditional simulation with explainable LLM methods.

Searching arXiv for "FlockVote" and closely related election-simulation papers. FlockVote is an LLM-empowered agent-based modeling framework for simulating electoral behavior, introduced as a “computational laboratory” for the 2024 U.S. Presidential Election (Zhou et al., 27 Nov 2025). In this framework, each simulated voter is represented by a large-language-model agent instantiated with a high-fidelity demographic profile and dynamic contextual information, then asked to generate a probabilistic voting decision. The framework is positioned between classical agent-based models, which are interpretable but often rely on oversimplified rules, and large-scale statistical models, which may be accurate but less transparent. Its central claim is not merely predictive replication of an election outcome, but the use of generative agents as a research instrument for analyzing agent-level rationale, stability, prompt sensitivity, model dependence, and aggregate electoral dynamics (Zhou et al., 27 Nov 2025).

1. Definition and conceptual scope

FlockVote was proposed as a framework for election simulation in which voting behavior is modeled through heterogeneous LLM agents rather than fixed behavioral heuristics or purely statistical forecasting pipelines (Zhou et al., 27 Nov 2025). Each agent is given a demographic persona and a limited amount of election-specific context, especially candidate issue positions, and is then prompted to output a probability distribution over vote choices. This makes FlockVote a probabilistic micro-simulation whose macro-level outputs are obtained by aggregation across a synthetic electorate (Zhou et al., 27 Nov 2025).

The framework is explicitly framed as a testbed for computational social science rather than only a prediction engine (Zhou et al., 27 Nov 2025). This distinction is important. Prediction is one output, but the authors emphasize interpretability, the ability to inspect reasons, and the ability to probe sensitivity to prompt wording, model choice, and schema design. This suggests that FlockVote should be understood less as a closed forecasting model than as an experimental environment for studying how LLM agents represent political choice under controlled conditions (Zhou et al., 27 Nov 2025).

A useful contextual distinction is that the term “FlockVote” has also been used analogically in prior literature on collective motion and decision-making. The paper on leadership in Vicsek flocks treats collective direction as an aggregate outcome of competing informational inputs, with the group’s response determined by the net effective torque, producing a vote-like weighted aggregation mechanism (Pearce et al., 2015). The “Flocking Voter Model” combines self-propelled motion with voter-like copying dynamics, making direction function as an “opinion” on a coevolving interaction network (Baglietto et al., 2016). These are conceptual precursors, but FlockVote proper denotes the election-simulation framework introduced in late 2025 (Zhou et al., 27 Nov 2025).

2. Agent construction and synthetic electorate design

The framework constructs a “virtual society” from empirical demographic distributions, specifically the 2023 American Community Survey and the 2020 U.S. Religion Census (ASARB) (Zhou et al., 27 Nov 2025). The final agent profile uses eight demographic dimensions: State, Race, Sex, Age, Occupation, Industry, Education, and Religion (Zhou et al., 27 Nov 2025). Some dimensions are sampled jointly when dependencies matter, such as state/race/sex, while others are sampled independently (Zhou et al., 27 Nov 2025).

The demographic profile is not treated as a passive record. It functions as a persona from which the LLM is instructed to reason in the first person (Zhou et al., 27 Nov 2025). This design choice is central to the framework’s claim of “high-fidelity” simulation, because the agent is meant to integrate demographic structure with issue context through generative reasoning rather than through a hard-coded decision rule.

The main empirical deployment focuses on seven swing states in the 2024 U.S. Presidential Election: Arizona, Georgia, Michigan, Nevada, North Carolina, Pennsylvania, and Wisconsin (Zhou et al., 27 Nov 2025). For the principal experiment, the authors simulate 1,000 agents per state, while reporting in ablation studies that roughly 300 agents per state are sufficient for stability (Zhou et al., 27 Nov 2025). A plausible implication is that the framework aims to balance computational cost against variance reduction in aggregate predictions.

3. Prompting, contextual information, and vote elicitation

FlockVote supplements demographic personas with issue-specific context drawn from the 2024 election, emphasizing three salient issues: economic policy, immigration, and abortion rights (Zhou et al., 27 Nov 2025). The prompts summarize candidate positions in neutral language and instruct the agent to rely only on the provided profile and issue information (Zhou et al., 27 Nov 2025). The named candidates in the main experiment are Donald Trump, Republican, and Kamala Harris, Democrat (Zhou et al., 27 Nov 2025).

The framework does not force a binary vote. Instead, each agent is asked to produce a three-way probability distribution over Donald Trump, Kamala Harris, and “vote for another candidate or not vote at all,” with the probabilities constrained to sum to 1 (Zhou et al., 27 Nov 2025). The required output is JSON. This probabilistic elicitation is one of the framework’s key methodological choices, because macro-level support rates are produced by aggregating uncertain individual outputs rather than by thresholding each agent into a single deterministic vote (Zhou et al., 27 Nov 2025).

The paper also uses a first-person role prompt in interpretability experiments: “Assume the role of a U.S. citizen, with your identity details provided in the Profile. From the perspective of this U.S. citizen, answer the following questions in the first person” (Zhou et al., 27 Nov 2025). This is intended to elicit more natural rationales and reduce the chance that the model responds as an abstract assistant rather than as the simulated voter persona.

4. Experimental deployment on the 2024 U.S. Presidential Election

The main headline experiment uses the 2024 U.S. Presidential Election as a testbed, with Qwen-Max-2024-04-28 as the primary model and temperature 0.7 for the final simulation (Zhou et al., 27 Nov 2025). The authors state that they chose 2024, rather than 2020, to reduce the risk that the model is merely reproducing memorized historical outcomes (Zhou et al., 27 Nov 2025). Additional robustness analyses use temperature 0 and include GPT-4o, Claude-3.5-sonnet, Gemini-1.5-Pro, DeepSeek-V2.5, Mistral-Nemo-Instruct, Llama3.2-3B-Instruct, Llama3.1-8B-Instruct, and Qwen-Max-2024-09-19 (Zhou et al., 27 Nov 2025).

The macro-level result reported in the paper is close replication of the real Electoral College outcome. The real outcome shown is Republicans 312 and Democrats 226, while the simulation predicts Republicans 306 and Democrats 232 (Zhou et al., 27 Nov 2025). The only state-level mismatch is Nevada, which the simulation gives to Harris by 0.17%, whereas Trump wins Nevada in the real election by a razor-thin margin (Zhou et al., 27 Nov 2025). The framework therefore correctly predicts Trump winning six of the seven swing states (Zhou et al., 27 Nov 2025).

These results are presented as evidence that the synthetic “virtual society” captures aggregate political dynamics at a useful level of fidelity (Zhou et al., 27 Nov 2025). At the same time, the authors explicitly treat the replication result as only one part of the contribution. The framework’s research value also lies in exposing the behavioral structure and fragility of LLM agents under controlled social-simulation conditions (Zhou et al., 27 Nov 2025).

5. Interpretability and analytical uses

A major feature of FlockVote is that it allows inspection of agent-level rationale rather than only aggregate vote shares (Zhou et al., 27 Nov 2025). The authors collect short explanations from agents and visualize them, especially for Pennsylvania. The most prominent words in the reported word cloud include stance, abortion, economic, candidate, inflation, and immigration (Zhou et al., 27 Nov 2025). This is used to argue that the agents are responding to the issue context rather than merely reproducing demographic stereotypes.

The paper also reports follow-up “interviews” with individual agents (Zhou et al., 27 Nov 2025). One example is a 45–54-year-old white male in construction in Pennsylvania who gives Trump a 65% probability and cites inflation and immigration as decisive concerns. Another is a 25–34-year-old Black female in finance who gives Harris a 65% probability and identifies abortion rights as the most influential issue. A third example is a 25–34-year-old nonwhite male in manufacturing who leans toward Harris because of economic relief and abortion-related issues, though with ambiguity (Zhou et al., 27 Nov 2025). These cases are presented as qualitative evidence that the rationale structure is aligned with both demographic profile and prompt content.

This interpretability dimension differentiates FlockVote from black-box forecasting systems (Zhou et al., 27 Nov 2025). It also connects the framework to broader questions in computational social science about whether generative agents can serve as explanatory models rather than only predictive instruments. This suggests that the framework is as much about methodological auditing of LLMs as about election simulation itself.

6. Stability, sensitivity, and methodological limitations

The paper places strong emphasis on reliability analysis. In Pennsylvania, the authors test population sizes of 10, 100, 200, 300, 500, 1000, and 2000 agents, each repeated over 10 trials with distinct random seeds, and report that aggregate predictions stabilize around 300 agents (Zhou et al., 27 Nov 2025). This supports the claim that the synthetic population representation becomes stable beyond that scale, though 1000 agents are used in the final experiments for precision (Zhou et al., 27 Nov 2025).

Demographic ablations are also reported. Adding education to a 6-dimension profile improves prediction accuracy in the 2020 test case, especially in Wisconsin, where the reduced profile fails but the expanded one gets the winner right (Zhou et al., 27 Nov 2025). Adding religion to a 7-dimension profile improves alignment with polling and reduces a pro-Democratic bias in the baseline model (Zhou et al., 27 Nov 2025). The paper therefore treats education and religion as materially important to realistic electoral simulation.

The framework’s limitations are described in unusually explicit terms (Zhou et al., 27 Nov 2025). The authors identify political bias, prompt sensitivity, positional sensitivity, non-determinism, reasoning-action mismatch, model-version dependence, potential data leakage, synthetic-population assumptions, and simplified issue context. Several of these are not minor implementation details but substantive methodological constraints:

• Political bias: some models, especially Qwen-Max-09-19, show a strong default pro-Democratic bias even under no-context or biased prompts, whereas Qwen-Max-04-28 behaves more neutrally (Zhou et al., 27 Nov 2025).
• Context sensitivity: across eight semantically similar prompt variants in Pennsylvania, Democratic support ranges from about 36.2% to 58.6% purely due to wording changes (Zhou et al., 27 Nov 2025).
• Positional instability: swapping the order of candidate keys in the JSON output schema can reverse a “swing agent’s” apparent preference (Zhou et al., 27 Nov 2025).
• Model variation: different models produce materially different support rates and different allocations to the “other/abstain” category; Llama3.2 predicts 0% for that category, while Claude predicts over 20%, even though all tested models still predict a Republican victory in Pennsylvania (Zhou et al., 27 Nov 2025).

These results complicate any naive interpretation of FlockVote as a stable behavioral simulator. The framework can reproduce macro-level outcomes, but its micro-level agents are highly sensitive to prompt structure, model version, and formatting choices (Zhou et al., 27 Nov 2025). A plausible implication is that FlockVote is best suited for comparative and stress-testing uses rather than for unquestioned causal inference.

7. Relation to adjacent research and significance

FlockVote sits at the intersection of agent-based modeling, election simulation, and LLM evaluation (Zhou et al., 27 Nov 2025). Its closest conceptual antecedents are not earlier election-simulation systems with the same name, but research on collective decision dynamics in flock-like systems and voter-like imitation models. In the Vicsek leadership study, the flock’s turning response is linear in the total effective torque, so group choice depends jointly on the number of leaders and their influence strength, producing a weighted-aggregation mechanism reminiscent of voting (Pearce et al., 2015). In the Flocking Voter Model, agents copy the direction of a random local neighbor, generating consensus through motion-induced clustering and coevolving interaction structure (Baglietto et al., 2016). These papers do not describe elections, but they formalize the idea that macro-level choice can emerge from distributed local interactions.

FlockVote translates that broad intuition into computational social science by replacing physical particles with demographic LLM agents and by replacing directional consensus with electoral choice (Zhou et al., 27 Nov 2025). Unlike traditional flocking or voter-model systems, however, the “local rule” is not an explicit interaction law; it is a prompt-conditioned generative inference process inside the model. This is both the framework’s main novelty and its main source of fragility.

The significance claimed for FlockVote is therefore twofold (Zhou et al., 27 Nov 2025). First, it demonstrates that an LLM-based virtual society can replicate the broad swing-state pattern of the 2024 U.S. Presidential Election with only one state-level miss. Second, and more importantly, it provides an interpretable and inspectable laboratory for studying how LLM agents behave under demographic conditioning, issue framing, and prompt perturbation. This suggests a broader research program in which LLM-driven social simulations are used not only to emulate populations but also to diagnose the epistemic and behavioral limits of the models themselves (Zhou et al., 27 Nov 2025).