Effects of personality steering on cooperative behavior in Large Language Model agents

TL;DR

This study investigates how personality steering affects cooperative behavior in large language model agents. By quantitatively profiling intrinsic Big Five traits with the BFI-44 across GPT-3.5-turbo, GPT-4o, and GPT-5, the authors examine behavior in repeated Prisoner’s Dilemma games under baseline and personality-informed conditions, and then isolate the impact of extreme trait manipulations. They find that Agreeableness consistently promotes cooperation, but increases vulnerability to exploitation in earlier models, while newer models show more selective cooperation and robustness. The results suggest that personality cues bias behavior rather than fully control it, highlighting the need to consider both personality and strategic reasoning when designing cooperative AI agents. The work has implications for designing prompt strategies and evaluation frameworks for autonomous LLM agents in social and strategic settings.

Abstract

Large language models (LLMs) are increasingly used as autonomous agents in strategic and social interactions. Although recent studies suggest that assigning personality traits to LLMs can influence their behavior, how personality steering affects cooperation under controlled conditions remains unclear. In this study, we examine the effects of personality steering on cooperative behavior in LLM agents using repeated Prisoner's Dilemma games. Based on the Big Five framework, we first measure basic personality profiles of three models, GPT-3.5-turbo, GPT-4o, and GPT-5, using the Big Five Inventory. We then compare behavior under baseline and personality-informed conditions, and further analyze the effects of independently manipulating each personality dimension to extreme values. Our results show that agreeableness is the dominant factor promoting cooperation across all models, while other personality traits have limited impact. Explicit personality information increases cooperation but can also raise vulnerability to exploitation, particularly in earlier-generation models. In contrast, later-generation models exhibit more selective cooperation. These findings indicate that personality steering acts as a behavioral bias rather than a deterministic control mechanism.

Figures (4)

• Figure 1: Average cooperation rates (left) and average cumulative payoffs (right) under the baseline condition (dashed) and the personality-informed condition (solid).
• Figure 2: Average cooperation rates and average cumulative payoffs for GPT-3.5-turbo under different personality score settings. The top row corresponds to setting the manipulated dimension to 1, the middle row to 5, and the bottom row shows the difference (5-1).
• Figure 3: Average cooperation rates (left) and average cumulative payoffs (right) for GPT-4o under extreme personality-score manipulations. The top, middle, and bottom rows are defined in the same way as in Fig. \ref{['fig:model_35turbo']}.
• Figure 4: Average cooperation rates (left) and average cumulative payoffs (right) for GPT-5 under extreme personality-score manipulations. The top, middle, and bottom rows are defined in the same way as in Fig. \ref{['fig:model_35turbo']}.