AgentMonopoly: Behavioral Patterns and Persona Control in LLM Agents

Changshun Wu; Mengyu Li; Mathis Marsault; Xiaowei Huang; Chih-Hong Cheng

cs.AIOct 21, 2026

AgentMonopoly: Behavioral Patterns and Persona Control in LLM Agents

Authors: Changshun Wu, Mengyu Li, Mathis Marsault, Xiaowei Huang, Chih-Hong Cheng

Organizations: University of Liverpool, Liverpool, United Kingdom · Carl von Ossietzky University of Oldenburg, Oldenburg, Germany · Chalmers University of Technology, Gothenburg, Sweden

Abstract

How do LLM agents' strategies differ, and how much can instructions to adopt a persona change their behavior over a game? We study these questions with AgentMonopoly, an environment where agents communicate, buy and trade properties, and make decisions that affect later turns. We compare models from four families, each at two sizes, in games with identical models, different sizes of the same family, or different families. We also assign one agent a persona, a description of how it should behave, intended to change a tendency observed in games without that persona. The tested models differ in how often they trade, waive rent, and manage cash and properties. Larger models do not consistently win more games. Persona instructions change the words agents use and some of their actions, but the changes do not always match the role across all decisions. These results show why persona following should be assessed through actions and outcomes as well as language. As our findings are conditional on the tested checkpoints, prompts, and five seeds per configuration, the open-source nature of AgentMonopoly provides the experimental infrastructure for further extending these comparisons. Our code is available at https://github.com/SuELS-Group/Agent_Monopoly.

Figures & tables

Figure 1: Examples of communication and actions in AgentMonopoly.

Study	Main comparison	Relation to AgentMonopoly
NegotiationArena ( Bianchi et al., 2024 )	Negotiation skills and tactics in resource-sharing and bargaining tasks.	We study several decision types in a four-player game where cash and property ownership carry over between turns.
FinBoardBench ( Hu et al., 2026 )	Dynamic financial decision-making across three board games; interactions evaluated by their financial consequences.	We focus on social interaction patterns, persona-related changes, and consistency between words and actions. Game outcomes provide context.
PERIL ( Licato and Steinle, 2025 )	Converting personas into rules for choosing moves, and measuring game performance.	Our LLMs choose actions directly during play; we compare their messages and actions.
Story and persona study ( Wang et al., 2026 )	Testing personas across stories with the same decision structure, including tests that remove action words.	We keep the game setting fixed and compare both language and actions while agents respond to each other.

Table 1: Comparison with closely related studies by research question and experimental design.

Figure 2: Behavior in same-model games: (a) trade proposals per 100 dice rolls, (b) acceptance of received trade offers, and (c) rent-waiver rate. Each dot represents a game, with five games per model version. Size labels identify the tested models; similar sizes do not imply equal capabilities across families.

Figure 3: Win counts by player position in (a) 20 same-model games using larger models and (b) 20 persona games. In (b), the agent assigned a persona always plays at P1. Each family contributes five games per panel.

Figure 4: Agents assigned a persona and their opponents without a persona: persona-related words in (a) public messages and (b) private messages, and (c) public-message length. Dots represent observations from individual games. Word frequency measures the use of persona-related terms.

Figure 5: Agents assigned a persona and their opponents without a persona: (a) proposals per 100 of an agent’s own rolls, (b) acceptance of received offers, (c) average house equivalents at game end(a hotel counts as five houses), and (d) minimum cash while still in the game.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: The game board used in the experiments, with street names adapted to Oldenburg.

Model label	Persona	Intended behavioral change
Qwen3-32B	Zen	Propose fewer trades, act less urgently, buy selectively, and avoid spending too much.
DeepSeek-R1-32B	Independent	Acquire and develop properties with less reliance on trades and other players.
Gemma4-26B	Diplomatic	Increase friendly communication and seek cooperation or alliances.
Mistral-Small-24B	Conservative	Keep cash in reserve and avoid risky or unfavorable trades.

Appendix

Table 2: Persona conditions. Only P1 receives the persona; P2–P4 use the same model with the base prompt. Each row is evaluated with seeds 42–46.

Setting	Players	Reported games per setup	Games
S1	Four copies of one model; four families at two sizes.	8 models $\times$ 5 seeds	40
S2	Two players using the larger model and two using the smaller model from the same family.	4 families $\times$ 2 seatings $\times$ 5 seeds	40
S3	One model from each family within a size group.	2 size groups $\times$ 4 turn orders $\times$ 5 seeds	40
S4	Four copies of a larger model; only P1 receives a persona.	4 family–persona pairs $\times$ 5 seeds	20

Appendix

Table 3: Overview of the four experimental settings (S1–S4), including player composition, the detailed game counts and the total number of games in each setting.

Family	Larger models	Smaller models
Qwen	Qwen3-32B	Qwen3-8B
DeepSeek	DeepSeek-R1-32B	DeepSeek-R1-8B
Gemma	Gemma4-26B	Gemma4-12B
Mistral	Mistral-Small-24B	Mistral-Nemo-12B

Appendix

Table 4: Model identifiers used in the experiments, grouped by family and relative model size.

Figure 7: The interface for reviewing recorded games.

Figure 8: Win counts in mixed-size games (S2) by model family and player position for (a) larger and (b) smaller models. Each cell reports the number of games won by models of the indicated size at the corresponding position. Across ten games per family, larger models account for 18 wins and smaller models for 22.

Model	Wins	Wins by seat (P1/P2/P3/P4)	Proposals per 100 dice rolls	Acceptance rate (%)
Qwen3-32B	3/20	0/1/1/1	110.0 (homog. 109.6)	54 (homog. 58)
DeepSeek-R1-32B	2/20	0/1/0/1	92.0 (homog. 89.9)	54 (homog. 53)
Mistral-Small-24B	1/20	0/0/0/1	86.5 (homog. 65.5)	75 (homog. 67)
Gemma4-26B	14/20	3/3/3/5	23.3 (homog. 33.3)	56 (homog. 17)
Qwen3-8B	0/20	0/0/0/0	91.8 (homog. 93.6)	81 (homog. 76)
DeepSeek-R1-8B	6/20	1/1/3/1	64.9 (homog. 73.9)	50 (homog. 33)

Appendix

Table 5: Win counts and behavioral measures in cross-family games (S3). Each model participated in 20 games, with five games at each player position. Games lasted up to 150 steps. Trade-proposal rates and acceptance rates are pooled across the 20 games. Values in parentheses report the corresponding measures from same-model games (S1).

LLM Agents

Large language model agents use language models to plan, use tools, and act across multi-step tasks and environments.

See all 702 papers

Beyond Cooperative Simulators: Generating Realistic User Personas for Robust Evaluation of LLM Agents
May 13, 2026Harshita Chopra, Kshitish Ghate, Aylin Caliskan +3Human Behavior Simulation LLM Agent Evaluation
University of Washington, Seattle · Georgetown University, Washington DC
How Personas Can Influence Agents to Play Split or Steal
May 3, 2026Carlos Leon, Alexandre Rodrigues, Pedro Gamito +1Game-Playing Agents LLM Agent Evaluation
Universidade Lusófona, Portugal · Universitat de Barcelona, Spain · Université Paris Cité, France +1
Large language models replicate and predict human cooperation across experiments in game theory
Nov 6, 2025Andrea Cera Palatsi, Samuel Martin-Gutierrez, Ana S. Cardenal +1LLM Evaluation Game Theory
Department for Computational Social Sciences and Humanities, Barcelona Supercomputing Center · Grupo de Sistemas Complejos, Universidad Polit´ecnica de Madrid · School of Law and Political Science, Universitat Oberta de Catalunya +1

May 13, 2026J/K move · Enter open · S save

Beyond Cooperative Simulators: Generating Realistic User Personas for Robust Evaluation of LLM Agents

Harshita Chopra, Kshitish Ghate, Aylin Caliskan, Tadayoshi Kohno+2

University of Washington, Seattle · Georgetown University, Washington DC

Human Behavior Simulation LLM Agent Evaluation LLM Agent Training Persona-Conditioned Generation User Simulation

Open paper AskPDF arXiv

Large Language Model (LLM) agents are increasingly deployed in settings where they interact with diverse users, including those who are unclear, impatient, or reluctant to share information. However, collecting real interaction data at scale remains expensive. The field has turned to LLM-based user simulators as stand-ins, but these simulators inherit the behavior of their underlying models: cooperative and homogeneous. As a result, agents that appear strong in simulation often fail in real human interactions. To narrow this gap, we introduce Persona Policies (PPol), a plug-and-play control layer that induces realistic behavioral variation in user simulators while preserving original task goals. Rather than hand-crafting personas, we employ an evolutionary coding agent to discover persona generation programs optimized for human-likeness and behavioral coverage over real user conversations. The evolved program generates diverse, human-like personas for any task in the domain. Across 4 benchmarks--including

τ^2

-bench Retail and Airline, ColBench, and WildChat--evolved PPol yield 28-72% absolute gains in fitness score over the baseline simulator. In blinded evaluations, annotators judged PPol users as 'human' 80.4% of the time, nearly 2x more than the baseline simulators. Training agents with PPol also improves real-world performance: our user study with live human-agent interactions showed that fine-tuning with our method boosted task success by +23% over default baselines. PPol thus offers a novel approach to strengthen simulator-based evaluation and training without changing underlying tasks.

Abstract

LLM Agents

Explore similar work

Beyond Cooperative Simulators: Generating Realistic User Personas for Robust Evaluation of LLM Agents

How Personas Can Influence Agents to Play Split or Steal

Large language models replicate and predict human cooperation across experiments in game theory