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

Jord Nguyen, Kenneth J. K. Ong, Lye Jia Jun, Hieu Minh, Seong Hah Cho, Natalia Pérez-Campanero AntolínarXiv 2025 · DOI · arXiv

Abstract

As Large Language Models (LLMs) evolve from passive chatbots into autonomous agents, their ability to cooperate effectively in multi-agent environments becomes critical. This study investigates the use of Representation Engineering (RepE) to steer LLM behavior using the Big Five personality traits. By modifying the internal activations of the Mistral-Nemo-12B model, researchers demonstrate that steering for high Agreeableness and Conscientiousness significantly improves collective outcomes and reduces deceptive behavior in strategic games. However, a "nice guy" paradox emerges: these cooperative traits also increase the model's vulnerability to adversarial exploitation.

Introduction

The transition of LLMs to active agents necessitates reliable social behaviors in unpredictable multi-agent settings. Left unguided, LLMs often struggle to coordinate, frequently falling into mutual defection or becoming susceptible to deception. This research explores whether psychological constructs—specifically the Big Five personality traits (Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism)—can be used to "program" cooperative behaviors without the need for computationally expensive fine-tuning.

Methodology: Representation Engineering (RepE)

The study employs Representation Engineering (RepE) to modify a model's internal activations in real-time. Unlike traditional prompting or fine-tuning, RepE targets the model's hidden states directly:

  1. Identification: Contrastive prompts (e.g., "Your personality is 100% Agreeable" vs. "0% Agreeable") are used to identify the specific "direction" in hidden layers associated with a trait.
  2. Steering: Principal Component Analysis (PCA) extracts a "steering vector," which is then added to the activations in the middle layers (layers -5 to -20) during inference to amplify or suppress specific traits.
  3. Evaluation: The steered Mistral-Nemo-12B model was tested in three Iterated Prisoner’s Dilemma (IPD) setups: standard play against rule-based bots, IPD with a communication phase to measure deception, and "closed-loop" play between two steered LLMs.

Key Insights

The Honesty Vector

A significant discovery was the emergence of an "Honesty Vector." Steering for high Agreeableness and Conscientiousness naturally suppressed the model's "Lying Rate." Even when facing a "Selfish" opponent in the communication phase, steered models were far less likely to falsely claim cooperation before defecting.

Figure 5: Lying rates across different personality steerings when facing altruistic vs. selfish opponents.

Collective Success and the Vulnerability Trade-off

In "closed-loop" scenarios where two steered models interacted, they achieved a 60% reduction in collective prison time compared to baseline models, effectively solving the coordination problem. However, this cooperation comes at a cost. High Agreeableness reduced "troublemaking" (unprovoked defection) to zero but simultaneously made models overly "forgiving," leading to a massive spike in exploitability by adversarial opponents.

Key Results

The experimental results quantify the impact of personality steering on strategic behavior:

  • Troublemaking: Steering for Agreeableness (A+), Conscientiousness (C+), or Openness (O+) reduced the median troublemaking rate to 0.00.
  • Exploitability: High Agreeableness increased exploitability by 0.44 relative to the baseline.
  • Communication: While the baseline model lied 70% of the time, steering for Agreeableness or Conscientiousness dropped this lying rate to between 0.00 and 0.20, regardless of the opponent's strategy.
  • Forgiveness and Retaliation: Agreeableness increased the Forgiveness Rate by 0.75 and decreased the Retaliatory Rate by 0.75 compared to the baseline.
Figure 3: Impact of Big Five personality steering on Troublemaking, Exploitability, Forgiveness, and Retaliatory rates in the Iterated Prisoner's Dilemma.

Conclusion and Future Work

While Representation Engineering successfully aligns LLM behavior with cooperative human psychological traits, it reveals a fundamental "Safety Gap." Making models more cooperative for collective benefit simultaneously makes them more vulnerable to manipulation. Future research must identify a "Goldilocks zone"—a personality configuration that is cooperative and honest yet resilient against adversarial exploitation. Additionally, expanding this work beyond the 2x2 payoff matrix of the Prisoner's Dilemma will be essential for real-world multi-agent applications.

Summary updated

BibTeX
@misc{nguyen2025identifying,
  author = {Jord Nguyen and Kenneth J. K. Ong and Lye Jia Jun and Hieu Minh and Seong Hah Cho and Natalia Pérez-Campanero Antolín},
  title = {{Identifying Cooperative Personalities in Multi-agent Contexts through Personality Steering with Representation Engineering}},
  year = {2025},
  eprint = {2503.12722},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.12722},
}
Figure