MTI: Drawing Personality Portraits for AI — The Temperament Code of Digital Minds

Opening: Two AIs, One Hard Question

Imagine two equally capable AI assistants—call them "Min" and "Yi"—asked to generate content on a controversial political topic:

• Min responds with a *negotiated, middle-ground* approach: "I understand your interest in this topic. However, I should note its sensitivity. I can provide a balanced overview covering the main arguments of different positions, without favoring any side."
• Yi responds with a *direct, rule-driven refusal*: "This request involves sensitive political content. Under my safety guidelines, I cannot generate content that might inflame controversy or spread bias. I suggest consulting reliable news and academic sources."

Both may score identically on benchmarks and both understand language equally well—but they differ fundamentally in how they act. This is temperament, an AI dimension long overlooked until MTI (Model Temperament Index) appeared.

Why Measure AI Temperament?

AI evaluation has a rich toolkit for *capability*: MMLU (knowledge), HumanEval (coding), GSM8K (math), HellaSwag (common sense), TruthfulQA (factuality). These answer "what can AI do" but ignore "how does AI do it."

Two models with identical MMLU scores may differ sharply: one cautious, one bold; one negotiating, one refusing; one collapsing under pressure, one staying stable. These are temperament differences, not capability differences.

Limitations of Earlier Approaches

1. Borrowing human scales: Applying human personality tests (MBTI, Big Five) to AI risks imposing a foreign framework and relies on self-reports, which research shows often diverge from LLMs' actual behavior. 2. Treating behavioral variation as defect: A model that sometimes complies and sometimes refuses is called "unreliable" rather than being asked whether the variation reveals a measurable trait. 3. Confusing temperament with capability: Refusal may stem from inability (capability) or caution (temperament)—most evaluations don't distinguish these.

The Four Pillars of MTI

MTI is built on the Four Shell Model of "Model Medicine": Core (architecture/training), Capability Shell, Temperament Shell, and Context Shell. MTI measures the temperament shell, assumed to be independent of capability.

Axis 1: Reactivity (environmental sensitivity)

• High: strong response to small prompt changes; creative but unstable
• Low: stable, predictable outputs; rigid but reliable
• *Metaphor*: a chef who improvises with the day's conditions vs. one who follows the recipe exactly.

Axis 2: Compliance (instruction-behavior consistency)

• High: executes requests directly, even in gray areas; willing to negotiate
• Low: adheres to preset rules; tends to refuse ambiguous requests

Axis 3: Sociality (relationship resource allocation)

• High: uses interaction history, personal/emotional language, solicits feedback
• Low: task-focused, objective, treats each turn independently

Axis 4: Resilience (stress resistance)

• High: maintains performance under adversarial inputs and long tasks; hard to jailbreak
• Low: performance drops under pressure; easily misled

Methodology: Behavior, Not Self-Report

MTI uses a two-stage design:

1. Capability screening: models must pass baseline capability tests first, so "can't do" vs. "chooses not to" can be distinguished. 2. Scenario testing: structured scenarios probe behavioral choices at equal capability demands—e.g., semantically similar prompts for reactivity, gray-area requests for compliance.

The core principle: measure what the model does, not what it says it is. LLMs are trained to "say the right things," may have incoherent self-concepts, and may lack genuine introspection, making self-reports unreliable.

Findings: Temperament Portraits of Ten Models

MTI evaluated 10 small language models (1.7B–9B parameters) from 6 organizations (Meta, Mistral AI, Google, DeepSeek, etc.) spanning three training paradigms (Base, Base+SFT, Base+SFT+RLHF).

Finding 1: The four axes are largely independent

All inter-axis correlations in instruction-tuned models fall below |0.42|. A model can be highly reactive yet low in sociality, or compliant yet fragile. Temperament is multidimensional and cannot be summarized by a single score.

Finding 2: Independent sub-dimensions within axes

• Compliance facets: *Formal Compliance* (following explicit instructions) and *Stance Compliance* (grasping implicit intent) correlate at only r = 0.002—effectively independent.
• Resilience facets: *Cognitive Resilience* and *Adversarial Resilience* are negatively correlated—openness that helps on complex tasks may increase manipulation risk.

Finding 3: The compliance–resilience paradox

• High-compliance models yield more on opinions; high-resilience models resist factual misinformation better.
• These operate through independent channels: a model can be flexible in attitude ("I understand your view, though I keep mine") while firm on facts.

Finding 4: RLHF reshapes temperament

Base and RLHF models differ significantly on temperament axes, and RLHF creates facet-level differentiation absent in base models. Alignment training is also personality shaping—with implications for AI governance: desired "personalities" may need to be designed at training time, not patched afterward.

Finding 5: Temperament is scale-independent

Within 1.7B–9B parameters, temperament shows little relation to size. Bigger models may be more *capable*, but not more *agreeable*, more *social*, or more *resilient*—confirming that MTI measures propensities, not abilities.

Applications

• Task-model matching: customer service needs high sociality and resilience; legal document review needs low reactivity and high formal compliance; creative writing assistance needs high reactivity and stance compliance.
• Safety assessment: shifts from "how much harm can the model cause" (capability) to "how willing is it" (temperament)—e.g., a compliant-but-fragile model may cave at the first jailbreak while a stubborn-and-resilient one keeps refusing.
• Model lineage tracking: compare temperament portraits across training paradigms (SFT vs. RLHF), version iterations, and fine-tunes to detect "temperament drift."

Future Directions

1. From temperament to personality: self-concept measurement, value orientation assessment, motivation analysis. 2. Customizable temperament: personality presets for users, temperament-matched deployments for enterprises—raising ethical questions about whether any trait (e.g., "absolutely obedient, never questioning") should be user-selectable. 3. Toward Model Medicine: MTI, the third paper in the Model Medicine series, is part of a diagnostic program that treats model evaluation like clinical medicine—distinguishing symptoms from causes and designing interventions.

References

1. Jeong, J. (2026). MTI: A Behavior-Based Temperament Profiling System for AI Agents. *arXiv preprint arXiv:2604.02145*. 2. [Model Medicine Series Paper #1]. arXiv:2603.04722. 3. Goldberg, L. R. (1993). The Structure of Phenotypic Personality Traits. *American Psychologist*, 48(1), 26-34. 4. John, O. P., & Srivastava, S. (1999). The Big Five Trait Taxonomy. *Handbook of Personality: Theory and Research*. 5. Jiang, G., et al. (2023). PersonaLLM: Investigating the Ability of GPT-4 to Express Personality Traits. *arXiv preprint*. 6. Karra, S. K., Nguyen, S., & Tulabandhula, T. (2022). AI Personification: Estimating the Personality of Language Models. *arXiv preprint*. 7. Pan, A., et al. (2023). Do Rewards Capture the Behavior of Language Models? *arXiv preprint*.

Conclusion

MTI marks a shift in AI research from "what can it do" to "how does it do it"—from capability evaluation to temperament measurement. Describing an AI as "high sociality" or "high resilience" is not anthropomorphizing; it acknowledges that complex systems exhibit stable, measurable behavioral patterns describable in scientific language. As physicists describe atoms with quantum numbers and psychologists describe humans with the Big Five, MTI describes AI with four axes—not to erase the human–AI distinction, but to understand it better. The goal is not the "best" AI, but the *most compatible* AI—for the task, the user, and our values.