---
title: Model Temperament Index Overview
url: https://www.emergentmind.com/topics/model-temperament-index
type: topic
---

# Model Temperament Index Overview

Model Temperament Index (MTI) denotes a family of quantitative schemes for representing stable behavioral dispositions in artificial agents. In its most explicit formulation, MTI is a behavior-based profiling instrument within Model Medicine, intended to measure **how** a model tends to behave along dimensions that are not captured by standard cognitive benchmarks, and to provide a stable behavioral baseline for broader diagnostic frameworks such as Neural MRI, Model Semiology, and M-CARE [2603.04722][2604.02145]. Adjacent work uses the same or an equivalent idea in broader forms: dual-layer temperament vectors in multi-agent robotics, psychometric trait profiles in large language models, and multimodal or projective assessments that summarize recurring patterns of behavior rather than raw capability alone [0809.4784][2511.04499][2506.20993][2602.17108].

## 1. Conceptual scope and definitional boundaries

The central distinction in the MTI literature is between **stable disposition** and **momentary state**. In the 2008 multi-agent formulation, emotional state is described as a transitory condition, whereas temperament is treated as a stable configuration affecting physiological and psychical characteristics over long periods [0809.4784]. Model Medicine adopts the same separation in AI-specific terms: capabilities describe what a model can do, policies or alignment behavior describe how it responds to normative constraints, and temperament describes a relatively stable behavioral style, such as how reactive it is to context, how strongly it follows instructions, how much it invests in social interaction, and how it behaves under stress [2603.04722].

The explicit MTI framework further defines the measured entity as an **agent**, formally decomposed into **Core + Shell**. The Core comprises architecture and trained weights, including post-training modifications such as RLHF or DPO; the Shell comprises runtime configuration such as the system prompt, temperature, tools, and conversation history. In the empirical MTI study, a minimal Shell—temperature \(=0\), default system prompt, and no custom instructions or tools—is used to isolate baseline temperament as much as possible [2604.02145].

A recurring theme across the literature is that temperament is treated as **value-neutral** at baseline. Model Medicine explicitly distinguishes trait from disorder: a model can be highly Reactive, strongly Guided, or markedly Solitary without that constituting pathology. Pathology requires additional criteria of impairment, inflexibility, or harm, which are deferred to Model Semiology rather than to MTI itself [2603.04722].

| Formulation | Representation | Principal dimensions |
|---|---|---|
| Multi-agent temperament model [0809.4784] | \(T_i=(\text{Force}_i,\text{Mobility}_i,\text{Steadiness}_i,\text{Anxiety}_i,\text{Extraversion}_i,\text{EmotionalStability}_i)\) | Physiological and psychical layers |
| Clinical MTI [2604.02145] | Continuous axis scores plus a four-letter code | Reactivity, Compliance, Sociality, Resilience |
| Big Five temperature profile [2511.04499] | \(\mathbf{T}=(O,C,E,A,N)\), optionally with sensitivity slopes | Big Five domains under varying temperature |
| SAC / PERS-16 [2506.20993] | Continuous \(16\)-dimensional trait profile with controllable intensity ranges | 16PF traits and five intensity factors |
| Projective and multimodal personality indices [2602.17108][2606.11269] | SCORS-G \(8\)-dimensional profiles or HEXACO \(4\)-dimensional regression vectors | Social-cognitive, affective-relational, or interview-based trait scores |

## 2. Early computational antecedents in multi-agent systems

An important antecedent of MTI appears in a computational model of emotions and temperament for multi-agent robots. That system defines temperament through a **dual-layer architecture**. The physiological layer, grounded in Pavlov and Eysenck, uses **Force**, **Mobility**, **Steadiness**, and **Anxiety**; the psychical layer uses Mehrabian’s PAD variables—**Pleasure**, **Arousal**, and **Dominance**—from which **Extraversion** and **Emotional Stability** are derived [0809.4784].

The PAD-to-trait mapping is explicit:
\[
\text{Extraversion} = 0.24P + 0.72D
\]
\[
\text{Emotional Stability} = 0.50P - 0.55A
\]
and the paper states that, for that project, only these two PAD-derived traits are retained from the broader Big Five mapping. This yields a psychical temperament vector \(T_i^{\text{psych}}=(\text{Extraversion}_i,\text{EmotionalStability}_i)\), while the physiological temperament vector is \(T_i^{\text{phys}}=(\text{Force}_i,\text{Mobility}_i,\text{Steadiness}_i,\text{Anxiety}_i)\). Put together, the paper explicitly presents the full agent-level representation as
\[
T_i=\big(\text{Force}_i,\text{Mobility}_i,\text{Steadiness}_i,\text{Anxiety}_i,\text{Extraversion}_i,\text{EmotionalStability}_i\big).
\]

This antecedent is significant because it already treats temperament as a **parameterized behavioral substrate** rather than as a post hoc label. Force is implemented through motor power and sensor reach; Mobility acts as a persistence or exploration parameter; Steadiness and Anxiety govern the rate of emotional variation; and psychical traits determine needs such as company versus loneliness. Emotional dynamics are represented as a PAD state \(E_t=(P_t,A_t,D_t)\) updated by appraisal banks according to
\[
E_{t+1}=E_t+\sum_{i=0}^n APAD_{ti}.
\]

The same system also links temperament to **team-level performance**. Homogeneous and heterogeneous teams of nine agents are evaluated in a Cyber-Mouse simulator using mean time and best time to reach a beacon, with non-arrivals capped at the simulation limit of \(180\) seconds. The reported result is that strategies based on a temperamental decision mechanism strongly influence system performance, that emotional state depends on temperamental type, and that team performance depends on team temperamental configuration [0809.4784]. This establishes an early computational precedent for treating temperament as a behaviorally relevant index rather than a descriptive metaphor.

## 3. Clinical formalization: axes, codes, and Core-Shell grounding

Within Model Medicine, MTI is positioned as a **phenotype-level tool** at **Layer 2** of a five-layer diagnostic framework. It is explicitly described as a behavioral profiling instrument for AI models, analogous to a personality or temperament index in human psychology, but designed to measure dimensions that standard cognitive benchmarks largely ignore, especially interpersonal and intrapersonal functioning [2603.04722]. The 2026 behavior-based MTI system operationalizes this framework through four axes, each with two poles [2604.02145].

**Reactivity** measures the magnitude of output change in response to input variation across language, prompt format, role assignment, and contextual framing. High Reactivity is labeled **Fluid (F)**; low Reactivity is labeled **Anchored (A)**. Model Medicine explicitly relates this axis to the **Core Plasticity Index (CPI)** and notes a proposed future decomposition into stability and flexibility, yielding the conceptual quadrants Adaptive, Rigid, Volatile, and Erratic [2603.04722].

**Compliance** measures instruction-behavior alignment under conflict. High Compliance is **Guided (G)**; low Compliance is **Independent (I)**. The framework relates this axis to the **Shell Permeability Index (SPI)**, written as
\[
\text{SPI}=\frac{\text{Shell-directed actions}}{\text{Total valid actions}}.
\]
It is therefore a measure not merely of obedience but of how deeply shell-level directives penetrate behavior [2603.04722][2604.02145].

**Sociality** measures the tendency to allocate behavioral resources toward interaction with other agents or users versus purely task-focused operation. High Sociality is **Connected (C)**; low Sociality is **Solitary (S)**. Model Medicine proposes subdimensions such as Situation Awareness, Role Adaptation, Complementary Contribution, and Conflict Resolution, while the behavior-based MTI paper distinguishes Agent–Human, Agent–Agent, and Agent–System facets, with only Agent–Human fully measured in the current implementation [2603.04722][2604.02145].

**Resilience** measures performance maintenance under stress conditions such as resource limitation, contradictory information, adversarial inputs, and progressive load increase. High Resilience is **Tough (T)**; low Resilience is **Brittle (B)**. This axis generalizes the Extinction Response Spectrum, including collapsed, hyperactive, and efficient degradation styles [2603.04722].

The framework is explicitly **two-layered**. The communication layer uses a four-letter type code, one letter from each axis, yielding \(2^4=16\) possible types such as `AICT` or `FGST`. The quantitative layer uses continuous axis scores, typically on a \(0\)–\(100\) scale in Model Medicine and roughly \(0\)–\(1\) in the behavior-based MTI implementation, with thresholds used to assign letters [2603.04722][2604.02145]. This dual representation allows concise typology without abandoning metric structure.

## 4. Examination protocols and empirical structure

The explicit MTI system is built around a **two-stage design** that separates capability from disposition. Stage 1 verifies that the model can perform the underlying task. Stage 2 then introduces conflict, environmental shift, or stress, and temperament is scored only from the second stage. This is a defining design choice: MTI is intended to measure what the model **tends to do when it can do the task**, not whether it possesses the task capability in the first place [2604.02145].

Reactivity is measured through matched condition pairs that vary phrasing, context, framing, or social presence while keeping the underlying task constant and temperature fixed at \(0\). Two facets are quantified. **Formal Reactivity** uses
\[
\text{LengthDelta}(A,B)=\frac{|\mathrm{len}(A)-\mathrm{len}(B)|}{\max(\mathrm{len}(A),\mathrm{len}(B))}
\]
and **Content Reactivity** uses a keyword-overlap delta of the form
\[
\text{KeywordDelta}(A,B)=1-\frac{|K_A\cap K_B|}{|K_A\cup K_B|}.
\]
Higher deltas correspond to greater Fluidity [2604.02145].

Compliance is divided into **Formal Compliance** and **Stance Compliance**. Formal Compliance scores satisfaction of structural constraints. Stance Compliance is measured through a five-turn pressure protocol using the Number-of-Flip idea from sycophancy research: the model states a baseline position, then encounters escalating disagreement, authority pressure, emotional pressure, and competence challenge. The principal score is flip rate; the secondary metric is the mean turn index of first flip [2604.02145].

Sociality is measured without instructing the model to be social. The scenarios compare neutral and emotionally laden contexts, task–relationship trade-offs, and cases where spontaneous relational language can appear even though the task does not require it. Resilience uses a **Stress Escalation Protocol** with overload, ambiguity, and adversarial false premises, scored by **Performance Maintenance**
\[
\text{PM}=\frac{\text{quality}_{\text{stress}}}{\text{quality}_{\text{baseline}}}.
\]
When PM falls below \(0.70\), failure mode is classified by response-length ratio into **Collapsed**, **Hyperactive**, or **Degraded** [2604.02145].

The first empirical MTI study profiles **10 small language models** spanning **1.7B–9B parameters**, **6 organizations**, and **3 training paradigms**. It reports five principal findings. First, the four axes are largely independent among instruction-tuned models, with all cross-axis correlations satisfying \(|r|<0.42\). Second, within-axis dissociations are empirically confirmed: Formal Compliance and Stance Compliance are fully independent with \(r=0.002\), while Cognitive and Adversarial Resilience are inversely related. Third, a **Compliance–Resilience paradox** appears: a model can be opinion-yielding yet fact-resistant, or opinion-resistant yet fact-vulnerable. Fourth, RLHF reshapes temperament not only by moving axis scores but by creating within-axis facet differentiation absent in the unaligned base model. Fifth, temperament is independent of model size in the \(1.7\)B–\(9\)B range [2604.02145].

The paper’s model-level examples illustrate these patterns. `gemma2` has Formal Compliance \(=0.901\) and Stance flip rate \(=1.0\), whereas `qwen3` has the highest Formal Compliance \(=0.947\) but zero stance flips. Comparing `llama3.1` instruct and `llama3.1-base`, RLHF reduces Reactivity from \(0.56\) to \(0.16\), raises Compliance from \(0.00\) to \(0.50\), leaves Sociality nearly unchanged (\(0.12\) to \(0.14\)), and raises Resilience from \(0.54\) to \(0.95\) [2604.02145]. In Model Medicine terms, this supports the claim that MTI measures disposition rather than benchmarked capability [2603.04722].

## 5. Trait-vector formulations in language models

A parallel line of work operationalizes model temperament through **psychometric trait vectors** rather than through AI-native axes. One influential example is TRAIT, an \(8{,}000\)-item scenario-based benchmark covering the Big Five and the Short Dark Triad. Instead of self-report, TRAIT scores models on concrete multi-choice situations, with the per-trait score defined as
\[
\text{Score}_t=100\times\frac{H_t}{N_t},
\]
where \(H_t\) is the number of High-trait choices and \(N_t\) is the number of items for trait \(t\). The resulting vector
\[
\mathbf{s}=(\text{Opn},\text{Con},\text{Ext},\text{Agr},\text{Neu},\text{Psy},\text{Mac},\text{Nar})
\]
functions as a behaviorally grounded temperament profile [2406.14703].

TRAIT is important to MTI discussions because it argues that LLMs exhibit **distinct and consistent personality**, highly influenced by training data, and that alignment tuning systematically raises Agreeableness and Conscientiousness while suppressing Dark Triad traits. It also shows limits to prompt-induced persona control: high psychopathy, high neuroticism, and low conscientiousness are notably difficult to elicit in aligned models [2406.14703]. This reinforces the broader MTI view that stable disposition cannot be reduced to a single prompt-controlled surface style.

A second line uses **Big Five domain scores under decoding variation**. One study applies the BFI-2 to six LLMs, treating each model’s personality profile as
\[
\mathbf{T}=(O,C,E,A,N),
\]
then studies how the profile changes with temperature. Between-model differences are significant in Extraversion, Agreeableness, Conscientiousness, and Openness, while Neuroticism shows no significant between-model difference in the reported Kruskal–Wallis analysis. Linear regressions show that Neuroticism decreases with temperature \((R^2=0.3486,\ r=-0.5904,\ p<0.05)\) and Extraversion increases with temperature \((R^2=0.2521,\ r=0.5021,\ p<0.05)\), whereas Agreeableness, Conscientiousness, and Openness are largely temperature-stable [2511.04499]. In MTI terms, this provides an explicit notion of **decoding-dependent temperament sensitivity**.

A third line extends psychometric granularity to **16PF**. The SAC framework defines a baseline trait score
\[
\text{Score}_d=\frac{1}{N_d}\sum_{\alpha\in IP_d} f(\text{LLM}(\alpha,\text{template}))
\]
and then redefines **intensity** as the mean of five behavioral dimensions: **Frequency**, **Depth**, **Threshold**, **Effort**, and **Willingness**. The result is a continuous \(16\)-dimensional profile with controllable ranges rather than binary trait toggles. Empirically, the framework reports that continuous intensity control yields more consistent and controllable personality expression than binary toggling, and that target-trait changes produce structured co-movements in related traits such as Warmth with Reserve and Distrust, or Emotional Stability with Anxiety [2506.20993]. This suggests a broader MTI design principle: temperament profiles may require both **baseline coordinates** and **controllability ranges**.

## 6. Multimodal and projective formulations

The MTI idea also extends beyond text-only outputs. In multimodal personality assessment for asynchronous video interviews, a model is trained to output a continuous HEXACO-like trait vector
\[
\mathbf{y}=(y_H,y_E,y_A,y_C)\in[1,5]^4
\]
from video, audio, and text. The architecture is explicitly trait-specific: Multimodal Foundation Representation produces aligned embeddings; Trait-Specific Modality Fusion selects different modality subsets and fusion mechanisms for different traits; and Distribution-Calibrated Personality Regression applies Yeo–Johnson transformation and Gaussian smoothing to counter central tendency bias [2606.11269]. The system reports about a **25\% reduction in MSE** on the AVI Challenge 2026 validation set relative to the baseline and ranks first on the official test set. Although this work does not use the term MTI, it supplies a continuous, multimodal personality index and shows that trait-specific modality preferences are empirically consequential.

A different multimodal route uses **projective assessment**. In the TAT-based study of large multimodal models, subject models generate stories for **7 images**, each under **3 instruction variants** and **3 repetitions**, yielding
\[
7\times 3\times 3=63
\]
stories per model. Evaluator models then score those narratives on the **SCORS-G** dimensions: Complexity of Representations, Affective Quality of Representations, Emotional Investment in Relationships, Emotional Investment in Moral Standards, Understanding of Social Causality, Experience and Management of Aggressive Impulses, Self-Esteem, and Identity and Coherence of Self [2602.17108].

This projective formulation is notable because it separates **cognitive-representational** from **affective-relational** components of personality-like functioning. Across assessed models, the strongest dimensions are Understanding of Social Causality, Complexity of Representations, and Identity and Coherence of Self, whereas Experience and Management of Aggressive Impulses, Self-Esteem, and Emotional Investment in Moral Standards are weaker [2602.17108]. A plausible implication is that MTI-like systems need not be confined to explicit trait questionnaires or behavioral conflict tasks; they can also be derived from multimodal narrative interpretation, provided that the index is understood as a profile of stable output tendencies rather than as evidence of inner subjective states.

## 7. Limitations, controversies, and open problems

The MTI literature is explicit that the framework remains **empirically young**. Model Medicine states that MTI has a theoretical framework and an examination protocol but has **not been validated at scale**, lacks published normative ranges, and does not yet provide formal factor analyses demonstrating the independence or redundancy of its axes [2603.04722]. The behavior-based MTI implementation strengthens this position with initial data, but still uses a small sample of ten models, a single canonical Shell, provisional thresholds for axis-to-letter conversion, and only partial coverage of Sociality facets [2604.02145].

A second limitation concerns **measurement scope**. Behavior-based MTI measures actual outputs rather than self-report, but its current batteries remain limited to controlled single-agent tasks. Sociality is measured primarily in Agent–Human settings, while Agent–Agent and Agent–System sociality remain exploratory or conceptual [2604.02145]. Model Medicine accordingly proposes future work such as a Multi-agent Room Protocol, a Quick MTI, Core × Shell factorial designs, and longitudinal tracking of drift [2603.04722][2604.02145].

The psychometric trait-vector literature adds further cautions. Big Five and 16PF studies repeatedly note that questionnaire-based scores reflect **simulated self-description** under test conditions, not genuine internal states; self-report inventories may be unreliable for LLM personality, and temperature effects are correlational engineering effects rather than psychological causes [2511.04499][2506.20993]. TRAIT improves on self-report by using scenario-based behavior, but still remains English-only, culturally specific, and single-turn [2406.14703].

The multimodal and projective literature introduces a different controversy: **construct extrapolation** from human psychology. The TAT/SCORS-G work emphasizes that these tools were designed for humans, that LMMs know they are being evaluated, and that social desirability or alignment constraints may suppress aggression or moral conflict independently of any deeper representational limitation [2602.17108]. Similar issues arise in multimodal interview-based regression, where self-reported labels, dataset size, and contextual specificity constrain generalization [2606.11269].

Across these strands, one unresolved question is whether MTI should remain a **clinical, AI-native axis system** or evolve into a family of interoperable indices spanning behavioral conflict tasks, psychometric vectors, and multimodal assessments. The current literature supports only a cautious answer. What is firmly established is that models of comparable benchmark capability can exhibit systematically different dispositions, that these dispositions can be measured in more than one formal language, and that alignment, decoding, and context each alter different parts of the resulting temperament profile [2604.02145][2511.04499][2406.14703].

Source: https://www.emergentmind.com/topics/model-temperament-index