TLDR: Agentmandering is a new framework that uses large language model (LLM) agents to create fair electoral districts. It redefines redistricting as a turn-based negotiation between two LLM agents representing opposing political parties, inspired by the ‘Choose-and-Freeze’ game theory protocol. Agents alternate between selecting a preferred map and freezing a district, ensuring balanced strategic interaction. This method significantly reduces partisan bias and unfairness, achieving much lower variance than traditional computational approaches, and has proven robust across different LLMs and settings.
Redrawing electoral district boundaries, a process known as redistricting, is fundamental to how votes translate into political power in representative democracies. This periodic redrawing, typically done after each decennial census, aims to ensure fair representation by reflecting population changes. However, this crucial process is often manipulated for political gain through a practice called partisan gerrymandering. This involves intentionally designing district lines to favor one political party, often by ‘packing’ voters into a few districts to concentrate their influence or ‘cracking’ them across many districts to dilute their voting power.
Traditional computational methods for redistricting have focused on generating a vast number of legally valid districting plans. While these methods, such as Markov Chain Monte Carlo (MCMC) and Sequential Monte Carlo (SMC), can produce thousands of plausible alternatives, they often overlook the strategic dynamics involved in selecting a final plan. This oversight creates opportunities for partisan actors to cherry-pick maps that, while technically compliant, are politically advantageous. Simply meeting formal constraints doesn’t guarantee fairness when the selection process itself can be manipulated.
Introducing Agentmandering: A New Approach to Fair Redistricting
A new framework called Agentmandering proposes a novel solution by reimagining redistricting as a turn-based negotiation between two agents representing opposing political interests. Drawing inspiration from game-theoretic ideas, particularly the “Choose-and-Freeze” protocol, this method embeds strategic interaction directly into the redistricting process using large language model (LLM) agents. These agents alternate between selecting and freezing districts from a small set of candidate maps, gradually partitioning the state through constrained and interpretable choices.
The core of Agentmandering is a sequential game played between a Republican agent and a Democratic agent. Each agent is powered by an LLM (such as Gemini 2.5 Pro) and is prompted to act in alignment with its party’s goals, aiming to defend and expand its representation based on state-specific political profiles, including historical voting trends, demographic composition, and partisan geography.
How Agentmandering Works
Each round of the game involves two key actions:
- Choose: One agent selects a preferred districting plan from a small set of candidate maps generated over the current unpartitioned region of the state. The candidate generator is party-agnostic and ensures that each proposed plan satisfies population balance, contiguity, and other legal constraints.
- Freeze: The opposing agent then selects one district from the chosen plan to be permanently fixed. This chosen district is removed from the unassigned region, and the process recurses on the remaining territory.
This iterative process continues until the entire state has been partitioned into districts. The game structure ensures that no single agent can unilaterally control the full outcome; instead, the final map emerges through a series of constrained, adversarial decisions, maintaining a strategic balance throughout.
Demonstrated Fairness and Stability
Evaluations using post-2020 U.S. Census data across all states have shown that Agentmandering significantly reduces partisan bias and unfairness. It also achieves remarkably lower variance—two to three orders of magnitude lower—than standard baseline methods. This indicates not only greater fairness but also enhanced stability in the resulting districting plans, particularly in competitive swing states.
The framework’s LLM-based agents proved more effective in making politically strategic decisions compared to rule-based variants, highlighting the value of integrating LLM reasoning with game-theoretic mechanisms to simulate human-like political behavior. Furthermore, Agentmandering demonstrated robustness across various LLM choices (including models from different national and institutional contexts like OpenAI, Anthropic, Mistral, Deepseek, and Qwen), suggesting that the Choose-and-Freeze strategy provides sufficient structural guidance to ensure fairness regardless of the underlying model’s specific biases or capabilities. The framework is also robust to the choice of candidate map generator and the initial agent ordering.
Also Read:
- Ensuring Fair Question Representation in Digital Deliberations
- InsurAgent: AI-Powered Simulation for Flood Insurance Decisions
A Step Towards More Equitable Representation
Agentmandering represents a significant advancement in computational redistricting. By harnessing LLM agents to implement game-theoretic negotiation, it transforms an abstract fairness mechanism into a scalable and practical solution for real-world challenges. The resulting plans are procedurally transparent, strategically robust, and empirically fair across multiple metrics, offering a new computational lens for political science and bridging the gap between theoretical fairness principles and applied algorithmic decision-making. For more details, you can refer to the full research paper here.
