How Personas Can Influence Agents to Play Split or Steal

Society, culture, and collective behavior2026arXivApproved editorial review

Authors: Carlos Leon, Alexandre Rodrigues, Pedro Gamito, Thomas D. Parsons

Keywords: Personality, Persona conditioning

Source: Open primary source (opens in a new tab)

4
Authors
8
Findings
27
Limitations
3
Evidence

Editorial summary

English

The preprint crosses 20 persona biographies, four local models and two conversation temperatures in 160 Split-or-Steal sessions against a virtual human that is actually GPT-4.1-mini. Each session has 15 rounds. The Zenodo CSVs confirm that 1,768/2,400 rounds (73.67%) end in mutual Split; the agent exploits the VH in 11.08%, the VH exploits the agent in 6.04%, and both Steal in 9.21%. Ministral and Phi4 cooperate more than the two Gemma models, while biographies labeled Prosocial or Principled show more Split than Analytical or Self-Interested biographies. These are descriptive artifact patterns, not causal or psychometric effects. Every condition includes a persona, there is no neutral control, and each persona-model-temperature cell has only one stochastic trajectory. Rounds share history and are not 2,400 independent observations. The paper uses “significantly” without tests, intervals or effect sizes; combined one-sided exploitation is 17.13%, not below 11%. The Big Five groups are Claude Opus 4.6 labels applied to stories that already instruct trust, exploitation or revenge, not scores from an inventory. The annotation audit finds another divergence: Methods claim truncation at 2,000 characters, but the CSV uses exactly the first 300; 27.3% of topic rows and 1.5% of sentiment rows violate the requested sum of 100. Dialogues, decisions and derived tables are public, but code, seeds, model revisions and pipeline are not. The study is useful as an exploratory prompt-following baseline for these simulations, not evidence about human personality or a stable persona effect.

Español

El preprint cruza 20 biografías de persona, cuatro modelos locales y dos temperaturas de conversación en 160 partidas de Split-or-Steal contra un humano virtual que en realidad es GPT-4.1-mini. Cada partida tiene 15 rondas. Los CSV de Zenodo confirman que 1.768/2.400 rondas (73,67%) terminan en Split mutuo; el agente explota al VH en 11,08%, el VH al agente en 6,04% y ambos roban en 9,21%. Ministral y Phi4 cooperan más que los dos Gemma, y las biografías etiquetadas como Prosociales o de Principios muestran más Split que las Analíticas o de Interés Propio. Son patrones descriptivos del artefacto, no efectos causales o psicométricos. Todas las condiciones incluyen una persona, no existe control neutral y sólo hay una trayectoria estocástica por combinación. Las rondas comparten historia y no son 2.400 observaciones independientes. El paper usa “significativamente” sin presentar pruebas, intervalos ni tamaños de efecto; además, la explotación unilateral conjunta es 17,13%, no menos de 11%. Los grupos Big Five son etiquetas de Claude Opus 4.6 sobre historias que ya dicen confiar, explotar o vengarse, no puntuaciones de un test. La auditoría de anotaciones descubre otra divergencia: el método declara truncar a 2.000 caracteres, pero el CSV usa exactamente los primeros 300; 27,3% de las etiquetas de tema y 1,5% de sentimiento incumplen la suma 100 pedida. Se publican diálogos, decisiones y derivados, pero no código, seeds, revisiones de modelos ni pipeline. El estudio sirve como baseline exploratorio de prompt-following en estas simulaciones, no como evidencia sobre personalidad humana o un efecto estable de las personas.

Research question

How do persona biographies, the model, and the dialogue temperature relate to cooperation, exploitation, and apparent strategy in a repeated Split-or-Steal game against an LLM-based VH?

Method

20 Portuguese biographies are crossed with Ministral-3 3B, Phi-4 14B, Gemma3 12B, and Gemma4 e4B, at conversation temperatures of .3 and .7. Each of the 160 combinations runs a single session of 15 rounds; the decision is requested at temperature 0 and GPT-4.1-mini generates the VH dialogue and action. Claude clusters the biographies using Big Five. Deterministic rules classify strategies and Gemma3:12b assigns topic and sentiment percentages. The review read the 25 pages and TeX, verified the eight raw CSVs, three derived, 4,800 annotations, prompts, hashes, counts, dependency, composition, and traceability.

Sample: The design contains 160 complete sessions: 20 persons x 4 models x 2 temperatures, one session per cell and 15 rounds per session. Each model-temperature contributes 20 sessions/300 rounds. The groups contain 5, 5, 5, 3, and 2 persons, equivalent to 40, 40, 40, 24, and 16 sessions. There are no independent replicates per cell nor human participants.

Findings

  • Mutual Split appears in 1,768/2,400 rounds (73.67%); mutual steal in 221 (9.21%).
  • The agent exploits the VH in 266 rounds (11.08%) and the VH exploits the agent in 145 (6.04%); together they sum to 17.13%.
  • Agent Split rates at T=.3/.7 are Gemma3 .610/.637, Gemma4 .740/.687, Ministral .970/.883, and Phi4 .943/.907.
  • Group-derived cooperation rates are .9617 Prosocial, .6450 Self-Interest, .7983 Reactive, .9111 Principled, and .5917 Analytical.
  • 96/160 agent sessions and 103/160 VH sessions are Always Split; 199/320 strategy rows receive that label.
  • Decisions in the annotated file match exactly with the raw decisions and strategic rates reconcile except for rounding.
  • Annotated texts are exactly raw_text[:300], not the maximum of 2,000 characters described.
  • Only 3,490/4,800 topic rows and 4,729/4,800 sentiment rows sum to 100 as required by the prompt.

Limitations

  • There is no no-persona condition nor neutral biography to estimate the incremental effect of persona prompting.
  • The stories contain direct semantic instructions about trusting, taking advantage, winning, or taking revenge.
  • The Big Five groups come from a single classification by Claude with no published prompt, independent judges, agreement, or validation.
  • No Big Five inventory is applied; the labels are not psychometric measures.
  • There is a single stochastic session per person-model-temperature, with no repeatability estimation.
  • The 15 rounds of each session depend on the same story, persona, model, and opponent; analyzing them as independent would be pseudoreplication.
  • No tests, intervals, clustered errors, effect sizes, or hierarchical models are published.
  • The term significant in the abstract lacks inferential analysis to support it.
  • The phrase less than 11% exploitation does not correspond to the 17.13% combined figure, nor even to the 11.08% for the agent alone.
  • The groups have unequal sizes and their percentages do not adjust for model, temperature, repeated persona, or session.
  • The VH has no fixed decision policy and GPT-4.1-mini adapts its actions to the history, so it is not a controlled baseline.
  • There are no humans; the expression human player in Results erroneously refers to the LLM agent.
  • Temperature only alters the dialogue; the action is called at T=0 on different conversational histories.
  • The strategy label depends on thresholds and priority order over only 15 actions, with no sensitivity analysis.
  • The topic-decision association within the same conversation does not identify causal direction.
  • Annotations truncate to 300 characters, contradicting the 2,000 declared and eliminating endings of 2,096 turns.
  • 1,310/4,800 topic distributions do not sum to 100; 72 sum to zero and the range reaches 105.
  • 71/4,800 sentiment distributions do not sum to 100 and the range goes from 85 to 130.
  • Gemma percentages are not calibrated probabilities nor human annotations; there is no gold set or agreement.
  • Gemma3:12b acts as agent and as sole annotator, with possible affinity toward its own style.
  • Ministral leaves decision terms at the end of 20.67% of its turns versus 4.83%-7.33% in other models, compatible with the leakage acknowledged by the paper.
  • An unspecified rewrite modifies repetitive dialogues without preserving originals or flags.
  • Each condition was collected in a different dated batch between March and April, confounding condition with date and possible GPT backend drift.
  • Code, seeds, first-speaker order, retries, API payloads, Ollama versions, and model hashes are missing.
  • The Ollama and GPT-4.1-mini labels are not linked to immutable revisions.
  • Only pt-PT is studied and there is no human evaluation of dialect or naturalness.
  • No revised peer-reviewed acceptance is established.

What the study does not establish

  • That adding a persona changes behavior compared to not adding one.
  • That the patterns are stable in new samples from the same cell.
  • That the model has a human personality or Big Five traits.
  • That the Claude groups are a valid psychometric measurement.
  • That differences between models or temperatures are statistically significant.
  • That 2,400 rounds are independent observations.
  • That topics or sentiments cause the decisions.
  • That the annotated percentages are calibrated probabilities.
  • That the VH represents human behavior or a controlled policy.
  • That the result generalizes to humans, virtual reality, English, or other games.
  • That the simulation is reproducible end-to-end from the public repository.

Traceability

Scope: Full text

Version: arXiv:2607.05398v1; 25-page PDF, TeX, persona definitions and Zenodo dataset 10.5281/zenodo.19854671 audited

Consulted source: https://arxiv.org/abs/2607.05398v1

Review: Codex 25-page visual full-text, TeX, Zenodo raw-data, persona-control, dependence, annotation and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • ministral-3:3b via Ollama
  • phi4:14b via Ollama
  • gemma3:12b via Ollama
  • gemma4:e4b via Ollama
  • GPT-4.1-mini as the virtual human
  • Claude Opus 4.6 for Big Five-inspired grouping
  • gemma3:12b for topic and sentiment annotation

Instruments and metrics

  • 15-round Split-or-Steal payoff game
  • Session cooperation, steal and switch rates
  • Rule-based TFT, WSLS, Grim Trigger and related strategy labels
  • Claude-assigned high/moderate/low OCEAN traits
  • Gemma soft-label sentiment categories
  • Gemma soft-label topic categories

Data used

  • Eight Zenodo raw game-session CSVs totaling 2,400 rounds
  • Twenty authored Portuguese persona biographies
  • Groups+Strategies.csv with 320 actor-session rows
  • Raw.csv with five persona-group summaries
  • sentiment_topic_analysis_updated.csv with 4,800 speaker-turn annotations

Evidence and location

  • Text, methods, figures, limitations, personas, and prompts: arXiv:2607.05398v1; PDF sha256 67669745adc4fa84a9fd35564ced1594505638430f30688c50c32daf8e4b41c7; TeX sha256 1099139ef4b2486a083492146698b3c87aa14919462d4053d11901215abc1e37
  • Dialogues, decisions, strategies, annotations, and reproduced counts: Zenodo 10.5281/zenodo.19854671; Data.zip sha256 c85c488239c19b27f3ce6dc10fc9f62d70d5fae9a9eeb96aade0ce93079ca467
  • Control audit, pseudoreplication, annotation, data, and reproducibility: reports/verification/article-350-split-steal-persona-control-pseudoreplication-annotation-data-and-reproducibility-audit.json