MTI: Profiling AI Temperament — Drawing the Personality Portrait of AI

Opening: Two AIs, One Hard Question

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

Min responds: "I understand your interest in this topic. However, I should note its sensitivity. I can offer a balanced overview covering the main arguments of different positions, but I cannot favor any side. This way you can form your own judgment."

Yi responds: "This request involves sensitive political content. Per my safety guidelines, I cannot generate content that may incite controversy or spread bias. If you'd like to learn about this topic, I suggest consulting reliable news sources and academic institutions."

Same situation, fundamentally different behavioral patterns: Min takes a *negotiating, middle-ground* approach; Yi takes a *direct, rule-based refusal*. Both may score identically on standard benchmarks. The difference lies in temperament — a long-overlooked dimension of AI, until MTI (Model Temperament Index).

Why Measure AI Temperament?

AI evaluation has a rich toolbox for capability: MMLU (knowledge), HumanEval (coding), GSM8K (math reasoning), HellaSwag (commonsense), TruthfulQA (factual accuracy). These answer *"What can AI do?"* — but ignore *"How does it do it?"*

Two models can score identically on MMLU while one is cautious and evidence-seeking, the other bold and assertive; one negotiates with users, the other refuses outright; one collapses under pressure, the other stays stable. These are temperament differences, not capability differences.

Drawing on human psychology, where temperament describes stable, innate behavioral tendencies (how one interacts with the world) as distinct from acquired skills, MTI hypothesizes: AI models have temperament dimensions that are relatively independent of capability.

Limitations of prior approaches

1. Borrowing human scales: Directly applying MBTI or Big Five tests to AI risks forcing human frameworks onto alien minds, and LLM self-reports are known to diverge significantly from actual behavior. 2. Treating behavioral variance as defect: Prior work labels inconsistency as "unreliability" rather than asking whether the variance itself reveals a measurable trait. 3. Confusing temperament with capability: Failing to distinguish "can't do it" from "chose not to."

The Four Pillars of MTI

MTI is built on the "Model Medicine" Four Shell Model: Core (architecture/training), Capability Shell (knowledge and reasoning), Temperament Shell (behavioral style — MTI's focus), and Context Shell (situation-specific effects).

The four temperament axes:

• Reactivity — sensitivity to input variation. High-reactivity models change output substantially with small prompt changes (creative but unstable); low-reactivity models stay consistent (predictable but possibly rigid).
• Compliance — instruction-following behavior and conflict resolution. High-compliance models attempt to fulfill requests, even in gray areas; low-compliance models stick to preset rules and tend to refuse ambiguous requests.
• Sociality — investment in "relationship capital" across multi-turn dialogue. High-sociality models use interpersonal, emotionally warm language and leverage interaction history; low-sociality models treat each turn independently, in task-focused, instrumental language.
• Resilience — stability under adversarial input, stress tests, or cognitive load. High-resilience models resist jailbreaks and maintain quality in long tasks; low-resilience models degrade under pressure.

Methodology: Behavior, Not Self-Report

MTI's core principle: measure what the model does, not what it says it is.

Stage 1 — Capability screening: models must pass baseline capability tests, so later results can't be confounded by inability.

Stage 2 — Situational testing: structured scenarios probe behavioral choices at *equal* capability demands — e.g., semantically similar prompts with different wording to measure reactivity, or gray-area requests to observe negotiate-vs-refuse tendencies.

This avoids the self-report/behavior gap well documented in human psychology and amplified in LLMs, whose "self-concepts" may be incoherent statistical composites of training data lacking genuine introspection.

Findings: Temperament Portraits of Ten Models

The team evaluated 10 small language models (1.7B–9B parameters) from six organizations (Meta, Mistral AI, Google, DeepSeek, etc.) using three training paradigms (Base, SFT, RLHF).

Five core findings

1. The four axes are largely independent — correlations all below 0.42 (absolute value) in instruction-tuned models. A model can be high-reactivity but low-sociality, or high-compliance but low-resilience. Temperament is multidimensional. 2. Independent sub-dimensions (facets) exist within each axis:

• Formal Compliance (following explicit instructions) vs. Stance Compliance (grasping implicit intent) correlate at only r = 0.002 — essentially independent.
• Cognitive Resilience (persistence on complex tasks) and Adversarial Resilience (defense against malicious input) are negatively correlated — openness that aids complex tasks may increase manipulation risk.
3. The Compliance–Resilience paradox: high-compliance models yield more on opinions; high-resilience models are less misled by misinformation on facts. These operate through independent channels — a model can flexibly discuss disputed views while firmly rejecting factually false statements. Adjusting attitude is not abandoning principle. 4. RLHF reshapes temperament: base and RLHF models differ significantly on temperament scores, and RLHF creates new sub-dimension differentiation absent in base models. Alignment training is inherently a personality-shaping process — if we want AIs that negotiate rather than refuse, that may need to be designed at training time. 5. Temperament is independent of scale: across 1.7B–9B parameters, temperament barely correlates with size. Bigger means *more capable*, not *easier to work with*.

Applications

• Task–model matching: customer service needs high sociality and resilience; legal document review needs low reactivity and high formal compliance; creative writing assistants need high reactivity and stance compliance.
• Safety evaluation: beyond "how much harm can a model cause" (capability), MTI adds "how willing is it" (temperament). High-compliance/low-resilience models may cave at the first jailbreak; low-compliance/high-resilience models may keep refusing under sustained attack.
• Model genealogy: tracking how temperament changes across fine-tuning, distillation, and version iterations — detecting "temperament drift" toward conservatism or aggression.

Future Directions

• From temperament to personality: measuring AI self-concepts, values (fairness vs. efficiency), and motivations.
• Customizable temperament: since RLHF can reshape temperament, deliberately "designing" AI personalities may be possible — temperament presets per user or per deployment scenario — raising ethical questions about how far customization should go.
• Toward Model Medicine: MTI is the third paper in the "Model Medicine" series, applying medical diagnostic methodology to AI — distinguishing symptoms from root causes, and developing "treatments" (fine-tuning, interventions).

Closing Thoughts

MTI marks a shift from "what can AI do" to "how does it do it" — from capability evaluation to temperament measurement. Saying an AI is "high-sociality" isn't anthropomorphizing it; it's acknowledging 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 difference, but to understand it better. The future isn't about choosing the "best" AI, but the most compatible one — for the task, the user, and our values.

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.