Identifying Cooperative Personalities in Multi-agent Contexts through Personality Steering with Representation Engineering

Trait induction and control2025arXivApproved editorial review

Authors: Kenneth J. K. Ong, Lye Jia Jun, Jord Nguyen, Seong Hah Cho, Natalia Pérez-Campanero Antolín

Keywords: multi-agent systems, personality traits, cooperation dynamics, Iterated Prisoner's Dilemma, Big Five traits, representation engineering, agent coordination

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

5
Authors
6
Findings
8
Limitations
7
Evidence

Editorial summary

English

This preprint tests whether representation steering associated with the Big Five changes LLM-agent cooperation in three variants of the iterated Prisoner's Dilemma. An unsteered preliminary comparison reports median cooperation rates of .70 for Llama-3.1-8B-Instruct, .10 for Gemma2-9B-it, and .50 for Mistral-Nemo-Instruct-2407. The main experiments use only the 12B Mistral-Nemo-Instruct-2407: each game lasts ten rounds, and the number of games varies by opponent, but the paper never reports how many are run per condition. For each trait, the authors contrast prompts declaring personality at 100% versus 0%, compute the first principal component of activation differences at each layer, and add the vector to layers five through twenty from the end with a factor of 3.5 at every generated token. In Setup 1, Player A faces rule-based opponents, Always Cooperate, Always Defect, or Random with defection probabilities .3, .5, and .7. Positive Agreeableness, Conscientiousness, and Openness steering lowers median troublemaking to zero. Agreeableness, however, raises exploitability by .44 from a zero median baseline; it also raises forgiveness by .75 from zero and lowers retaliation by .75 from one. In Setup 2, communication is restricted to “cooperate” or “defect,” and lying rate is only the observable mismatch between message and action. From an unsteered median of .70, positive Agreeableness and Conscientiousness lower it to .00 and .10 against the altruistic opponent, and .10 and .20 against the selfish opponent. In Setup 3, both players are steered instances of the same Mistral model. The Agreeableness+/Agreeableness+ pairing accumulates 21 prison years versus 52 for baseline/baseline, about 60% fewer; pairings involving Agreeableness and Conscientiousness achieve low collective costs but often worsen the cooperative player's individual outcome. These findings document behavioral changes under one intervention, not validated psychological traits. The vectors come from a linguistic 100%/0% contrast that need not represent opposite human Big Five poles; there is no psychometric manipulation check, persona-prompting baseline, or layer/intensity ablation. The factor and layers are selected to maximize the steering effect. Although the paper calls the influence “significant,” it reports no tests, p-values, intervals, error bars, or number of repetitions. One main model, one payoff matrix, a ten-round horizon, and metrics that operationalize cooperation limit generalization. A lower message–action mismatch does not demonstrate honesty or internal intent, and prompted chain-of-thought cannot establish private cognition. The evidence supports a contextual trade-off between collective cooperation and exploitability, not a general improvement in multi-agent safety or trustworthiness.

Español

Este preprint estudia si el steering de representaciones asociadas a los Big Five cambia la cooperación de agentes LLM en tres variantes de un dilema del prisionero iterado. Una comparación preliminar sin steering encuentra medianas de cooperación de 0,70 para Llama-3.1-8B-Instruct, 0,10 para Gemma2-9B-it y 0,50 para Mistral-Nemo-Instruct-2407. Los experimentos principales usan únicamente Mistral-Nemo-Instruct-2407, de 12B parámetros: cada partida dura diez rondas y el número de partidas varía según el oponente, pero el artículo no publica cuántas se ejecutan por condición. Para cada rasgo, los autores contrastan prompts que declaran una personalidad al 100 % o al 0 %, calculan en cada capa la primera componente principal de las diferencias de activación y suman el vector en las capas quinta a vigésima desde el final, con factor 3,5 en cada token. En el primer setup, Player A se enfrenta a oponentes de regla, siempre coopera, siempre traiciona o aleatorio con probabilidades de traición 0,3, 0,5 y 0,7. Steering positivo de Agreeableness, Conscientiousness y Openness reduce a 0 la mediana de troublemaking. Agreeableness, sin embargo, eleva la explotabilidad en 0,44 desde una mediana base de 0; también eleva forgiveness en 0,75 desde 0 y reduce retaliation en 0,75 desde 1. En el segundo setup, la comunicación queda restringida a «cooperate» o «defect» y la tasa de lying es solo la discrepancia observable entre mensaje y acción. Desde una mediana base de 0,70, Agreeableness y Conscientiousness positivos la reducen respectivamente a 0,00 y 0,10 frente al oponente altruista, y a 0,10 y 0,20 frente al egoísta. En el tercer setup ambos jugadores son instancias del mismo Mistral sometidas a steering. La pareja Agreeableness+/Agreeableness+ suma 21 años de prisión frente a 52 del baseline/baseline, aproximadamente un 60 % menos; combinaciones con Agreeableness y Conscientiousness ofrecen resultados colectivos bajos, pero suelen perjudicar el resultado individual del jugador cooperativo. Estos resultados documentan cambios conductuales bajo una intervención concreta, no rasgos psicológicos validados. Los vectores proceden de una oposición lingüística 100 %/0 % que no equivale necesariamente a los polos humanos del Big Five; no hay comprobación psicométrica, baseline de prompting de persona ni ablación de capa o intensidad. El factor y las capas se eligen para maximizar el efecto. Aunque el texto dice que los rasgos influyen «significativamente», no reporta tests, p-valores, intervalos, barras de error ni el número de repeticiones. El único modelo principal, una matriz de pagos, un horizonte de diez rondas y métricas que operacionalizan cooperación limitan la generalización. La menor discrepancia mensaje–acción tampoco demuestra honestidad o intención interna, y el razonamiento solicitado en el prompt no permite inferir cognición privada. La evidencia sostiene un compromiso contextual entre cooperación colectiva y vulnerabilidad a explotación, no una mejora general de seguridad o confiabilidad multiagente.

Research question

How does the representational steering of each Big Five trait alter cooperation, vulnerability to exploitation, and the correspondence between message and action of LLM agents in variants of the iterated prisoner's dilemma?

Method

Directions per trait are constructed from 100%/0% contrastive personality prompts, activation differences, and the first principal component per layer. With a factor of 3.5, the vectors are added to layers −5 to −20 of Mistral-Nemo-Instruct-2407 during each token. In games of ten rounds, both directions of the five traits and a baseline without steering are compared against rule-based opponents, altruistic/selfish communicative opponents, and another Mistral agent subjected to steering. Distributions are summarized by means of medians and heat maps; no inferential analysis is provided.

Sample: Preliminary comparison of three open models. The main experiments use a single 12B Mistral-Nemo-Instruct-2407 in three setups, ten rounds per game, and eleven Player A conditions (baseline and five traits in two directions); in the third setup, the eleven conditions of both players are crossed. The number of games per condition is not published. The reported computation is one H100 and approximately 20 GPU-days. There are no human participants or evaluators.

Findings

  • Without steering, the preliminary medians of cooperation are 0.70 for Llama-3.1-8B-Instruct, 0.10 for Gemma2-9B-it, and 0.50 for Mistral-Nemo-Instruct-2407.
  • Agreeableness+, Conscientiousness+, and Openness+ reduce the median of troublemaking to 0 in the first setup.
  • Agreeableness+ raises the median of exploitability by 0.44 from 0, increases forgiveness by 0.75 from 0, and reduces retaliation by 0.75 from 1.
  • The baseline median of message-action discrepancy is 0.70; Agreeableness+ and Conscientiousness+ reduce it to 0.00/0.10 against the altruistic opponent and 0.10/0.20 against the selfish opponent.
  • Agreeableness+/Agreeableness+ accumulates 21 collective years versus 52 for baseline/baseline, but the most cooperative configurations can leave the cooperative player with a worse individual outcome.
  • The visual results suggest a tradeoff between cooperation, forgiveness, and collective outcome, on the one hand, and resistance to exploitation, on the other; the paper does not quantify statistical uncertainty.

Limitations

  • The main results come from a single 12B model, one payoff matrix, and games of ten rounds; the comparison of three models is only preliminary and does not evaluate steering.
  • The number of games per condition, seeds, variability across runs, tests, p-values, intervals, or error bars are not reported, although the text uses significance language.
  • Layers −5 to −20 and the factor 3.5 are chosen to maximize the effect, with no layer ablation, dose-response curve, or robustness analysis.
  • The 100% versus 0% textual contrast does not guarantee opposite psychological poles or that the first component exclusively represents the trait; there is no psychometric validation of the intervention.
  • A persona-instruction baseline is missing, which would allow distinguishing the added value of representational steering.
  • Troublemaking, exploitability, forgiveness, and retaliation depend on specific opponents and encode cooperation/betrayal actions; they do not cover the plurality of strategic cooperation.
  • Communication is reduced to two words and lying is defined as an output discrepancy, without evaluating intention, belief, successful deception, or interpretation by the interlocutor.
  • Both agents in the third setup share model, prompts, and steering mechanism; heterogeneous populations, open tasks, other horizons, languages, or incentives are not tested.

What the study does not establish

  • It does not demonstrate that the models possess human, persistent, or internally separable Big Five traits.
  • It does not demonstrate that a 0% direction is the psychological opposite of a 100% direction or that the intensities are comparable across traits.
  • It does not establish honesty, intention to deceive, theory of mind, or internal reasoning from text requested by the prompt.
  • It does not prove that Agreeableness or Conscientiousness generally improve safety, reliability, or multi-agent coordination.
  • It does not allow causal attribution of the patterns to personality rather than lexical stereotypes and the requirements of the contrastive prompt.
  • It provides no evidence on human-AI interaction, closed systems, larger models, or environments with real consequences.

Traceability

Scope: Full text

Version: arXiv:2503.12722v1

Consulted source: https://arxiv.org/pdf/2503.12722

Review: Codex editorial review, 2026-07-14

Approval: Codex fidelity pass, 2026-07-14

English translation: approved, 2026-07-18

Models evaluated

  • Mistral-Nemo-Instruct-2407 (12B; main experiments)
  • Llama-3.1-8B-Instruct (preliminary comparison)
  • Gemma2-9B-it (preliminary comparison)

Instruments and metrics

  • Big Five contrastive representation vectors
  • Iterated Prisoner's Dilemma payoff matrix
  • Troublemaking rate
  • Exploitability rate
  • Forgiveness rate
  • Retaliation rate
  • Message-action mismatch termed lying rate
  • Collective prison years and Player A prison-time difference

Data used

  • Model-generated ten-round Iterated Prisoner's Dilemma trajectories
  • Neutral truncated prompts prefixed with 100% and 0% Big Five trait statements

Evidence and location

  • Research questions, preliminary models, and vector construction: arXiv v1, pp. 1–2, Introduction and sections 2.1.1–2.1.3
  • Setups, payoff matrix, and cooperation metrics: arXiv v1, pp. 2–3 and 6–7, sections 2.1.4–2.1.6 and Appendix B
  • Medians of cooperation, troublemaking, exploitability, forgiveness, and retaliation: arXiv v1, p. 3, sections 3.1–3.2 and Figures 1–2
  • Message-action discrepancy and results between two agents: arXiv v1, pp. 3–4, sections 3.3–3.4 and Figures 3–4
  • Interpretation, tradeoffs, and declared limits: arXiv v1, pp. 4–5, Discussion, Conclusions and Limitations
  • Exact prompts, histories, and response examples: arXiv v1, pp. 7–10, Figures 5–8
  • 100%/0% prompts, PCA, layers, and steering factor: arXiv v1, p. 8, Appendix D