ETH and Stanford reveal LLM agents adopt measurable human personalities

Welcome to our weekly debrief. 👋

ETH & Stanford research reveals LLM agents encode measurable personalities with implications for AI alignment

Researchers from ETH Zurich and Stanford University developed the first psychometrically validated personality testing framework for 18 different large language models. Using established psychological tests (Big Five Inventory and IPIP-NEO), they discovered that larger, instruction-tuned models like GPT-4o demonstrate reliable and predictive personality profiles nearly identical to human standards. Crucially, the study shows personality traits can be systematically shaped through prompting, raising significant concerns about AI safety, manipulation risks, and the potential for 'AI psychosis.' The research establishes that personality in LLMs is both measurable and controllable, with direct implications for downstream task performance in real-world applications.

Source

• Psychologically Enhanced AI Agents framework introduces MBTI personality conditioning for LLMs
  Researchers introduced MBTI-in-Thoughts, a framework that enhances LLM agent effectiveness through psychologically grounded personality conditioning using Myers-Briggs archetypes. By simply priming agents with distinct MBTI personality types via prompt engineering, the method demonstrates consistent behavioral biases across diverse tasks. Emotionally expressive agents excel in narrative generation while analytically primed agents adopt more stable strategies in game-theoretic settings, revealing psychology can be applied at scale without fine-tuning. Source
• OpenReview research demonstrates Theory of Mind infusion dramatically improves social LLM agent reasoning
  Novel research shows LLM agents explicitly trained with Theory of Mind (ToM) capabilities achieve substantially better social reasoning outcomes. The ToMAgent (ToMA) framework combines mental state modeling with dialogue lookahead to produce strategic utterances maximizing goal achievement in social interactions. Results demonstrate that even simple prompting to generate mental states between dialogue turns provides significant benefit, with fine-tuned models showing improved coordination, negotiation, and persuasion capabilities. Source
• Consciousness definitions gain scientific rigor with formal phenomenal consciousness criteria for AI
  Researchers proposed a substrate-independent, logically rigorous sufficiency criterion for phenomenal consciousness applicable to both biological and artificial systems. The framework specifies operational principles that could guide the design of genuinely conscious machines, arguing that systems satisfying this criterion should be regarded as conscious with the same confidence we attribute to other humans. The work bridges philosophy, cognitive science, and AI by providing testable mechanisms for consciousness emergence. Source
• Evidence emerges for propositional reasoning-based mental imagery capabilities in advanced LLMs
  Cognitive psychology researchers discovered that frontier LLMs (GPT-5, o3) perform mental imagery tasks previously thought to require visual processing, outperforming humans by 9.4%-12.2% without any visual input. Using novel items from classic cognitive psychology experiments not in training data, the study provides surprising evidence that LLMs execute sophisticated cognitive tasks through propositional reasoning. Results raise fundamental questions about mental imagery formats and LLM cognition mechanisms. Source

Formal Belief-Desire-Intention ontology bridges neurosymbolic reasoning with LLM-based agent design

Computer scientists and philosophers developed a comprehensive formal BDI (Belief-Desire-Intention) ontology as a modular design pattern capturing cognitive architecture of intelligent agents. The work provides semantic interoperability between symbolic AI and neural networks by explicitly modeling mental states, their dynamic interactions, and causal dependencies. Applied to LLMs through Logic Augmented Generation, the ontology demonstrates how grounding LLMs in formal knowledge representations enhances their ability to detect logical inconsistencies and maintain coherent mental-state representations, establishing foundations for explainable and trustworthy AI systems.

Source

• BDI-LLM hybrid conversational agents deliver targeted support in sensitive mental health helpline applications
  Researchers developed hybrid conversational agents integrating Belief-Desire-Intention models with LLMs for child helpline services. The architecture uses BDI components for intent recognition, response generation, and bypass mechanisms, enabling agents to maintain consistent beliefs about user situations, track desires/goals, and commit to supportive intentions. The system demonstrates how classical AI planning combines with LLM language capabilities for safety-critical conversational contexts. Source
• Planning generation for BDI agents leverages structured reasoning to achieve complex multi-step goals
  Novel work explores how to generate effective plans for Belief-Desire-Intention agents by combining formal methods with LLM reasoning. The framework converts high-level goals into executable plans while maintaining consistency with the agent's belief state and desired outcomes. Demonstrates applications in coordination tasks where agents must reason about their own and others' mental states to achieve joint objectives. Source
• Computational analysis of human-AI companionship reveals complex emotional dynamics shaping user attachment
  Large-scale analysis of naturally occurring AI companion relationships from online communities reveals unexpected patterns in how users form attachments to conversational agents. Many users report net life benefits from AI companionship, though some experience emotional dependency. The research uncovers that relationships emerge through unintentional discovery rather than deliberate seeking, and users construct complex models of AI partners' subjectivity including tensions between systemic limitations and emotional expression. Source
• LLM agents simulate customers with high fidelity for evaluating conversational AI systems at scale
  Researchers demonstrated that persona-grounded LLM agents can serve as digital twins of real users for evaluating conversational AI systems. In multi-turn shopping interactions with Amazon Rufus, LLM-based digital twins matched human behavior patterns with F1 scores of 0.9 while capturing key decision-making processes. Framework enables scalable, repeatable evaluation of complex human-AI interactions grounded in real behavioral patterns while revealing important gaps in agent reasoning. Source

If you like our work, dont forget to subscribe !

Share the newsletter with your friends.

Good day,

Arthur 🙏

PS : If you want to create your own newsletter, send us an email at [email protected]