TLDR: A new framework, MBTI-in-Thoughts, uses prompt engineering to give Large Language Model (LLM) agents distinct personalities based on the Myers-Briggs Type Indicator (MBTI). This psychological priming allows control over AI behavior along cognitive and affective axes, leading to predictable biases. For instance, “feeling” agents excel in creative writing, while “thinking” agents adopt stable strategies in games. The framework also supports multi-agent communication, showing that self-reflection improves cooperation, and is generalizable to other personality models, all without fine-tuning.
Large Language Models (LLMs) are rapidly changing how we interact with technology, and their influence is growing beyond just basic functions. Researchers are now exploring how to give these AI models distinct personalities to make them even more effective and human-like.
A new framework called MBTI-in-Thoughts (MiT) introduces a novel way to enhance LLM agents by conditioning them with psychologically grounded personality traits. This method draws inspiration from the well-known Myers-Briggs Type Indicator (MBTI) and uses clever prompt engineering to imbue agents with specific personality archetypes. The goal is to control their behavior along two fundamental dimensions of human psychology: cognition (reasoning, planning) and affect (emotion, empathy).
The core idea is simple yet powerful: by giving an AI agent a personality, you can influence how it approaches different tasks. For example, the research shows that agents primed to be more ’emotional’ excel at generating creative narratives, producing stories that are more emotionally charged, personal, and optimistic. On the other hand, agents primed for ‘analytical’ thinking adopt more stable and consistent strategies in complex game-theoretic scenarios, like the Prisoner’s Dilemma.
How Does it Work?
The MiT framework has two main components. First, individual agents are ‘primed’ with a psychological profile using structured prompts. This involves giving the LLM a role-setting context and specific behavioral directives. To ensure this priming is effective and consistent, the framework integrates the official 16Personalities test for automated verification. This means the AI agent takes a personality test, and its responses are checked to confirm they align with its assigned MBTI profile.
Second, the framework enables structured multi-agent communication. This allows researchers to study how personality influences interactions between multiple AI agents. Different communication protocols are explored, from simple majority voting to more complex interactive dialogues with self-reflection. Interestingly, the study found that encouraging agents to ‘self-reflect’ privately before interacting with others significantly improves cooperation and the quality of their collective reasoning.
Personality in Action
The research highlights several fascinating behavioral biases induced by personality priming:
- Feeling vs. Thinking: ‘Feeling’ types generate more emotionally expressive narratives, while ‘Thinking’ types show more rigid but consistent strategies in adversarial games, often defecting more to maximize individual gain. This suggests a trade-off between adaptability and stable planning.
- Introversion vs. Extraversion: ‘Introverted’ agents exhibit higher truthfulness in communication and demonstrate more reflective internal cognition, producing longer and more elaborated rationales. ‘Extraverted’ agents, conversely, are associated with social risk-taking.
- Judging vs. Perceiving: ‘Judging’ agents tend to be more truthful than ‘Perceivers,’ aligning with their preference for structure, reliability, and rule-following.
These findings suggest that personality priming can be a lightweight yet effective mechanism to align AI agent traits with specific task demands. For instance, ‘Feeling’ or ‘Introverted’ profiles could be ideal for sensitive applications requiring empathy and trust, such as healthcare or negotiation. ‘Judging’ profiles might enhance structured planning, while ‘Perceiving’ profiles offer adaptability in rapidly changing environments.
Also Read:
- Beyond the Black Box: LLMs and the Evolution of AI Agents
- AI Agents Master Complex Tasks by Integrating Linguistic Guidance and Direct Experience
Beyond MBTI
While the focus is on MBTI, the researchers emphasize that their approach is highly generalizable. The framework can seamlessly integrate with other established psychological models like the Big Five (OCEAN), HEXACO, Enneagram, and DISC. This is because all these frameworks can be abstracted into a shared formal structure, where personality types are defined as configurations over continuous psychological dimensions.
This groundbreaking work, detailed in the research paper Psychologically Enhanced AI Agents, establishes a foundation for creating AI agents that are not only more capable but also socially aligned and trustworthy, opening new avenues for human-AI interaction and collaborative reasoning.
