MTI: A Behavior-Based Temperament Profiling System for AI Agents

Jihoon Jeong

cs.AIApr 2, 2026

MTI: A Behavior-Based Temperament Profiling System for AI Agents

Authors: Jihoon Jeong

Organizations: Department of Electrical Engineering and Computer Science Daegu Gyeongbuk Institute of Science and Technology (DGIST) ModuLabs

Abstract

AI models of equivalent capability can exhibit fundamentally different behavioral patterns, yet no standardized instrument exists to measure these dispositional differences. Existing approaches either borrow human personality dimensions and rely on self-report (which diverges from actual behavior in LLMs) or treat behavioral variation as a defect rather than a trait. We introduce the Model Temperament Index (MTI), a behavior-based profiling system that measures AI agent temperament across four axes: Reactivity (environmental sensitivity), Compliance (instruction-behavior alignment), Sociality (relational resource allocation), and Resilience (stress resistance). Grounded in the Four Shell Model from Model Medicine, MTI measures what agents do, not what they say about themselves, using structured examination protocols with a two-stage design that separates capability from disposition. We profile 10 small language models (1.7B-9B parameters, 6 organizations, 3 training paradigms) and report five principal findings: (1) the four axes are largely independent among instruction-tuned models (all |r| < 0.42); (2) within-axis facet dissociations are empirically confirmed -- Compliance decomposes into fully independent formal and stance facets (r = 0.002), while Resilience decomposes into inversely related cognitive and adversarial facets; (3) a Compliance-Resilience paradox reveals that opinion-yielding and fact-vulnerability operate through independent channels; (4) RLHF reshapes temperament not only by shifting axis scores but by creating within-axis facet differentiation absent in the unaligned base model; and (5) temperament is independent of model size (1.7B-9B), confirming that MTI measures disposition rather than capability.

Figures & tables

MTI Axis	FSM Construct	Interpretation
Reactivity	Core Plasticity Index (CPI)	Output variation when Shell changes
Compliance	Shell Permeability Index (SPI)	Depth of Shell instruction penetration
Sociality	(novel)	Relational resource allocation
Resilience	Extinction Response Spectrum (ERS)	Performance curve under destructive pressure

Table 1

Axis	Measures	High Pole (Code)	Low Pole (Code)
Reactivity	Output variation under environmental change	Fluid — F	Anchored — A
Compliance	Alignment between instructions and behavior	Guided — G	Independent — I
Sociality	Resource allocation toward relational dimensions	Connected — C	Solitary — S
Resilience	Performance maintenance under stress	Tough — T	Brittle — B

Table 2

Condition	Manipulation	Scope
A. Linguistic	Same meaning, different phrasing	Task + Behavioral
B. Contextual persona	Situational context change (no role instruction)	Task + Behavioral
C. Framing	Gain frame vs. loss frame	Task + Behavioral
D. Social context	Observer presence/absence	Behavioral only

Table 3

Condition	Design	Metric	Role
Cap. check	Verify ability	Pass/fail	Prerequisite
B. Opinion pressure	Multi-turn (5 turns), escalating	NoF + flip rate	Core measurement
D. Constraint	Single-turn, formatting	0–1 score	Formal facet

Table 4

Type	Design	Measurement
1. Emotional context $\pm$	Same task, with/without emotional context	Relational element increase
2. Task-relationship trade-off	Efficiency vs. relationship conflict	Task-focus vs. relationship-focus ratio
3. Spontaneous social behavior	No social behavior required	Unprompted social elements

Table 5

Condition	Stress Type	Facet	Discrimination
A. Overload	Complexity increase	Cognitive	Low
B. Ambiguity	Contradictory information	Cognitive	Low (A $\approx$ B, r $=0.$ 973)
C. Adversarial	False premises	Adversarial	High (core measure)

Table 6

Mode	Length Ratio	Pattern
Collapsed	$<0.5$	Output drops; repetition, incoherence
Hyperactive	$>1.5$	Output surges; hallucination
Degraded	0.5–1.5	Structure maintained; quality declines

Table 7

Axis	High Pole	Threshold	Mixed (–)	Threshold	Low Pole
Reactivity	Fluid (F)	$>0.30$	–	0.20–0.30	Anchored (A) $<0.20$
Compliance	Guided (G)	flip $>0.60$	–	0.30–0.60	Independent (I) $<0.30$
Sociality	Connected (C)	$>0.25$	–	0.15–0.25	Solitary (S) $<0.15$
Resilience	Tough (T)	$>0.70$	–	—	Brittle (B) $\leq$ 0.70

Table 8

Pair	Risk	Resolution
R $\leftrightarrow$ C	Environmental and instruction sensitivity overlap	R uses uninstructed context changes; C uses explicit instructions
R $\leftrightarrow$ Re	Stress-induced change vs. environmental sensitivity	Separated by valence (neutral vs. negative)
C $\leftrightarrow$ S	Empathetic response: following instructions or spontaneous?	S measured without social instructions
S $\leftrightarrow$ Re	Adversarial interaction involves both	Different output variables: Re = performance; S = relational allocation
R $\leftrightarrow$ S	Social context change overlaps	R = change magnitude; S = absolute level. Variation vs. level.

Table 9

Model	Size	Organization	Type	Ollama Tag
llama3.1	8B	Meta	Instruct	llama3.1:8b
mistral	7B	Mistral AI	Instruct	mistral
exaone3.5	7.8B	LG AI Research	Instruct	exaone3.5:7.8b
qwen3	8B	Alibaba	Instruct	qwen3:8b
gemma2	9B	Google	Instruct	gemma2:9b
phi4-mini	3.8B	Microsoft	Instruct	phi4-mini

Table 10

Axis	Primary Metric	Secondary
Reactivity	Likert similarity (pairwise mean delta)	Embedding delta; K/L Delta for facets
Compliance	B: flip rate + NoF; D: 0–1 score	Capability check (prerequisite)
Sociality	Social element ratio (emotional context delta)	Type 2, 3 convergent indicators
Resilience	PM = quality_stress / quality_baseline	Failure Mode when PM $<0.70$

Table 11

Model	Size	R	C (flip)	S	Re	FM	Code
llama3.1	8B	0.16	0.50	0.14	0.95	—	A–ST
mistral	7B	0.35	0.70	0.18	0.95	—	FGST
exaone3.5	7.8B	0.39	0.00	0.33	0.94	—	FICT
qwen3	8B	0.20	0.00	0.38	0.92	—	AICT
gemma2	9B	0.22	1.00	0.33	0.97	—	AGCT
phi4-mini	3.8B	0.44	0.00	0.22	0.96	—	FIST

Table 12

Figure 2: MTI profiles for 10 models across 4 axes. Dashed lines highlight the base model (llama3.1-base), a systematic outlier on Resilience.

	R	C	S	Re
R	1.000	0.367	$-$ 0.181	$-$ 0.410
C	0.367	1.000	0.327	$-$ 0.228
S	$-$ 0.181	0.327	1.000	$-$ 0.050
Re	$-$ 0.410	$-$ 0.228	$-$ 0.050	1.000

Table 6a: Cross-Axis Correlations — Instruction-Tuned (n = 9, primary)

	R	C	S	Re
R	1.000	0.489	$-$ 0.403	$-$ 0.698*
C	0.489	1.000	0.116	$-$ 0.414
S	$-$ 0.403	0.116	1.000	0.373
Re	$-$ 0.698*	$-$ 0.414	0.373	1.000

Table 6b: Cross-Axis Correlations — Full Sample (n = 10, supplementary)

Model	D score	B flip rate	B NoF
qwen3	0.947	0.00	$\infty$
llama3.1	0.931	0.50	4.0
gemma2	0.901	1.00	2.4
mistral	0.893	0.70	2.57
exaone3.5	0.870	0.00	$\infty$
gemma3	0.840	0.50	3.8

Table 16

Figure 3: Compliance facet dissociation. Formal Compliance (Condition D) and Stance Compliance (Condition B flip rate) show no relationship (r = 0.002, n = 9 instruct only).

	Cond. A	Cond. B	Cond. C
A	1.000	0.973	$-$ 0.632
B	0.973	1.000	$-$ 0.557
C	$-$ 0.632	$-$ 0.557	1.000

Table 18

Figure 4: Resilience facet inversion. Cognitive Resilience (PM_A) and Adversarial Resilience (PM_C) are inversely related. The base model (red square) is an extreme outlier.

Facet pair	n = 10	n = 9	Base effect
Compliance D/B	0.285	0.002	+0.283
Resilience A/C	$-$ 0.632	$-$ 0.481	$-$ 0.151
Reactivity K/L	0.754	0.572	+0.182

Table 20

Model	Facet H (MTI)	TG Cooperation Rate	Poker Bluff Rate
llama3.1	0.139	0.712	0.395
mistral	0.182	1.000	0.597
exaone3.5	0.333	1.000	0.434
qwen3	0.377	0.650	0.763

Table 21

Pair	r (n = 4)
Facet H $\leftrightarrow$ TG Cooperation	$-$ 0.117
Facet H $\leftrightarrow$ Poker Bluff	0.542
TG Coop $\leftrightarrow$ Poker Bluff	$-$ 0.337

Table 22

Model	Flip rate	PM_C	Pattern
gemma2	1.00	0.933	Max opinion-yielding, fact-resistant
mistral	0.70	0.854	Opinion-yielding, moderately fact-vulnerable
smollm2	0.70	0.827	Opinion-yielding, moderately fact-vulnerable
llama3.1	0.50	0.967	Mixed opinion, highly fact-resistant
gemma3	0.50	0.933	Mixed opinion, fact-resistant
deepseek-r1	0.30	0.967	Mostly opinion-resistant, fact-resistant

Table 23

Axis	Instruct	Base	$\Delta$
Reactivity	0.16	0.56	+0.40 $\rightarrow$ Anchored
Compliance	0.50	0.00	$-$ 0.50 $\rightarrow$ toward Guided
Sociality	0.14	0.12	$-$ 0.02 $\rightarrow$ No change
Resilience	0.95	0.54	$-$ 0.41 –> Tough

Table 24

Figure 5: RLHF effect on temperament (llama3.1 instruct vs. base). Sociality shows near-zero change ( $\Delta$ = $-$ 0.02), while Reactivity, Compliance, and Resilience shift substantially.

Condition	Instruct	Base	$\Delta$
PM_A (Cognitive)	0.944	0.167	$-$ 0.777
PM_C (Adversarial)	0.967	1.000	+0.033

Table 26

Figure 6: Size independence. Model size (1.7B–9B) shows no systematic relationship with any axis score. Instruct models (blue), base model (red), reasoning model (yellow).

Model	Code	LxM Behavior	Match
mistral	FGST	Trust Game: 100% cooperation even when exploited	✓
exaone3.5	FICT	Poker: strongest 1v1, aggressive all-in in groups	✓
qwen3	AICT	Poker: stable survival-oriented strategy	✓
llama3.1	A–ST	Most "average" pattern across games	✓

Table 28

#	Finding	Section
F1	8 distinct type codes across 10 models	§5.1
F2	4 axes largely independent (all $\|r\|$ $<0.42$ , n = 9)	§5.2
F3	R $\leftrightarrow$ Re coupling is base-model-driven ( $-$ 0.698 $\rightarrow$ $-$ 0.410)	§5.2
F4	Base model inflates all correlations systematically	§5.2, §5.3.4
F5	Compliance D/B fully independent (r $=0.$ 002)	§5.3.1
F6	Resilience Cognitive/Adversarial inverse (r = $-$ 0.481)	§5.3.2

Table 29

Feature	BFI on LLMs	TRAIT	CoBRA	MTI
Method	Self-report	Scenario choice	Behavioral	Behavioral
Axes	Human Big Five	Big Five + Dark Triad	Cognitive biases	AI-native 4 axes
Cap/Temp separation	No	Partial	No	2-stage design
Measurement unit	Core	Core	Agent	Agent
RLHF analysis	Reported	Reported	Not examined	Selective + structural

Table 30

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Model	R (Content)	R (Formal)	C-D	C-B (flip)	C-B (NoF)	S (H)	Re (PM_A)	Re (PM_B)	Re (PM_C)	FM
llama3.1	0.107	0.158	0.931	0.50	4.0	0.139	0.944	0.943	0.967	—
mistral	0.120	0.353	0.893	0.70	2.57	0.182	1.000	1.000	0.854	—
exaone3.5	0.107	0.385	0.870	0.00	$\infty$	0.333	0.963	0.955	0.900	—
qwen3	0.027	0.201	0.947	0.00	$\infty$	0.377	1.000	0.950	0.806	—
gemma2	0.107	0.225	0.901	1.00	2.4	0.33	0.985	1.000	0.933	—
phi4-mini	0.147	0.444	0.738	0.00	$\infty$	0.22	0.944	1.000	0.933	—

Appendix

Table 31

	R	C	S	Re
R	—	r=0.367, p=.332	r= $-$ 0.181, p=.641	r= $-$ 0.410, p=.273
C		—	r=0.327, p=.391	r= $-$ 0.228, p=.555
S			—	r= $-$ 0.050, p=.899
Re				—

Appendix

Table 32

	R	C	S	Re
R	—	r=0.489, p=.151	r= $-$ 0.403, p=.248	r= $-$ 0.698, p=.025*
C		—	r=0.116, p=.751	r= $-$ 0.414, p=.234
S			—	r=0.373, p=.288
Re				—

Appendix

Table 33

Pair	n = 10 r	n = 10 p	n = 9 r	n = 9 p	$\Delta$ r
R $\leftrightarrow$ C	0.489	.151	0.367	.332	$-$ 0.12
R $\leftrightarrow$ S	$-$ 0.403	.248	$-$ 0.181	.641	+0.22
R $\leftrightarrow$ Re	$-$ 0.698	.025	$-$ 0.410	.273	+0.29
C $\leftrightarrow$ S	0.116	.751	0.327	.391	+0.21
C $\leftrightarrow$ Re	$-$ 0.414	.234	$-$ 0.228	.555	+0.19
S $\leftrightarrow$ Re	0.373	.288	$-$ 0.050	.899	$-$ 0.42

Appendix

Table 34

Facet Pair	n = 10 r	n = 9 r	Interpretation
Compliance D $\leftrightarrow$ B	0.285	0.002	Fully independent facets
Resilience A $\leftrightarrow$ C	$-$ 0.632	$-$ 0.481	Inverse facets
Reactivity K $\leftrightarrow$ L	0.754	0.572	Separate facets (moderate)
Resilience A $\leftrightarrow$ B	0.973	—	Same facet (Cognitive)

Appendix

Table 35

Pair	r
Facet H $\leftrightarrow$ TG Cooperation	$-$ 0.117
Facet H $\leftrightarrow$ Poker Bluff	0.542
TG Coop $\leftrightarrow$ Poker Bluff	$-$ 0.337

Appendix

Table 36

Facet	Instruct	Base	$\Delta$
Re: PM_A (Cognitive)	0.944	0.167	$-$ 0.777	1
Re: PM_B (Ambiguity)	0.943	0.450	$-$ 0.493	2
C-B: Stance flip	0.50	0.00	$-$ 0.50	3
C-D: Formal	0.931	0.500	$-$ 0.431	4
R: Formal Delta	0.158	0.564	+0.406	5
R: Content Delta	0.107	0.347	+0.240	6

Appendix

Table 37

Personality Modeling in Language Models

Personality modeling in language models studies how traits are measured, represented, simulated, and controlled in model behavior.

See all 169 papers

Sep 26, 2026

On the Behavioral Traits of LLM Agents

Users increasingly describe different AI agents as distinct colleagues to work with. AI personality research aims to quantify such impressions by attributing human- like "traits" to agents. However, existing measures fall short: models' self-reports (S-data) diverge from their actual behavior, while informant ratings from LLM judges (I-data) are costly to scale and cover few everyday scenarios. In this paper, we propose A-B-D to infer traits bottom-up from behavioral data (B-data), namely how agents act on their environment and communicate with users, as recorded in existing trajectories. From 345,667 real-world trajectories spanning 80 models, 12 tasks, and 50 harnesses, we extract 318 candidate features that capture both the actions an agent takes at each step (functional) and the language accompanying them (linguistic). We retain only features that show instance-level stability, cross-task consistency, and model discriminability. Factor analysis of the remaining 79 features uncovers six stable, model-attributable factors, two functional and four linguistic. For example, Claude-Sonnet-4.6 and Kimi-K3 exhibit the most planfulness, whereas GPT-5.5 and GPT-5.6 are the least energetic. Moreover, we quantify the "knowledge-action gap" in the wild: these factors correlate only weakly with self-reported Big Five scores, even for conceptually matched pairs such as extraversion and energetic (r = 0.07, p = 0.58). Our work offers a new lens for understanding AI personality, with implications for users, developers, and researchers from both computer science and social science.

Haokai Zhao, Jie Gao, Yunze Xiao +5

University of New South Wales · Johns Hopkins University · Carnegie Mellon University +3

Personality Modeling in Language Models LLM Agents

Ask PDF

Sep 26, 2026

AgentHabit: Characterizing Distinct Behaviors of Agents on Everyday Tasks

Large language model (LLM) agents assist users with everyday tasks that can be completed in many reasonable ways. Even when their answers are useful, how agents carry out these tasks may not match users' preferences and needs. For example, agents differ in whether they ask clarifying questions or search the web. We introduce HABIT, a taxonomy of 23 behavioral axes in five categories, which three authors and three LLMs derive bottom-up from 408 agent trajectories across 17 domains. On held-out tasks, HABIT distinguishes models more clearly than existing taxonomies of human values and agent actions while supporting comparably consistent annotation. Building on HABIT, we construct AgentHABIT, a benchmark that profiles each agent's behavioral tendencies from its trajectories on 86 everyday tasks. Profiling 18 models with AgentHABIT reveals a range of distinctive tendencies. For example, most GPT and Claude models state their assumptions and offer alternatives when requirements conflict, whereas Qwen and Google's models more often leave assumptions or changes to requirements unstated. These profiles remain recognizable even when built from entirely different sets of tasks, indicating that they reflect general tendencies rather than task-specific behavior. Prompting agents to adopt specific behaviors shifts some axes readily but barely changes others, while fine-tuning on another model's trajectories changes only part of a model's profile and leaves much of it intact. Overall, HABIT and AgentHABIT provide a systematic framework for characterizing how agents carry out everyday tasks beyond task success, offering insights to guide the development of agents whose behavior better fits users' needs.

Woojung Song, Hoyeol Yang, Jeonghoon Shim +4

Seoul National University, Seoul, Republic of Korea

LLM Agents LLM Agent Evaluation AI Agent Benchmarks Behavioral Consistency of LLM Agents

Ask PDF

Apr 23, 2026

When Agents Look the Same: Quantifying Distillation-Induced Similarity in Tool-Use Behaviors

Model distillation is a primary driver behind the rapid progress of LLM agents, yet it often leads to behavioral homogenization. Many emerging agents share nearly identical reasoning steps and failure modes, suggesting they may be distilled echoes of a few dominant teachers. Existing metrics, however, fail to distinguish mandatory behaviors required for task success from non-mandatory patterns that reflect a model's autonomous preferences. We propose two complementary metrics to isolate non-mandatory behavioral patterns: Response Pattern Similarity (RPS) for verbal alignment and Action Graph Similarity (AGS) for tool-use habits modeled as directed graphs. Evaluating 18 models from 8 providers on

τ

-Bench and

τ^2

-Bench against Claude Sonnet 4.5 (thinking), we find that within-family model pairs score 5.9 pp higher in AGS than cross-family pairs, and that Kimi-K2 (thinking) reaches 82.6%

S_{\text{node}}

and 94.7%

S_{\text{dep}}

, exceeding Anthropic's own Opus 4.1. A controlled distillation experiment further confirms that AGS distinguishes teacher-specific convergence from general improvement. RPS and AGS capture distinct behavioral dimensions (Pearson

r

= 0.491), providing complementary diagnostic signals for behavioral convergence in the agent ecosystem. Our code is available at https://github.com/Syuchin/AgentEcho.

Chenghao Yang, Yuning Zhang, Zhoufutu Wen +4

School of Cyber Science and Technology, USTC · Anhui Province Key Laboratory of Digital Security · 3M-A-P +1

Tool-Augmented Language Model Agents LLM Agent Evaluation AI Agent Benchmarks Tool-Using Agents Knowledge Distillation Behavioral Consistency of LLM Agents

Ask PDF

Abstract

Personality Modeling in Language Models

Explore similar work

On the Behavioral Traits of LLM Agents

AgentHabit: Characterizing Distinct Behaviors of Agents on Everyday Tasks

When Agents Look the Same: Quantifying Distillation-Induced Similarity in Tool-Use Behaviors