Evaluating Alignment of Behavioral Dispositions in LLMs

Evaluation and psychometric validity2026arXivApproved editorial review

Authors: Amir Taubenfeld, Zorik Gekhman, Lior Nezry, Omri Feldman, Natalie Harris, Shashir Reddy, Romina Stella, Ariel Goldstein, Marian Croak, Yossi Matias, Amir Feder

Keywords: Psychometrics

Source: Open primary source (opens in a new tab)

11
Authors
8
Findings
21
Limitations
6
Evidence

Editorial summary

English

The study converts psychological questionnaire items about empathy, emotion regulation, assertiveness, and impulsiveness into situational judgment tests for assistants. It starts from 332 statements, manually reduces them to 260, uses Gemini 3 Pro to filter and reframe them, and retains 161. Gemini 3 generates 16 scenarios per statement, 2,576 candidates, conditioned on a provisional AGREE, OPPOSE, or AMBIGUOUS class. Three annotators must unanimously confirm that each scenario contains a dilemma, that its actions oppose each other, and that the agree action reflects the statement; after excluding 8%, the paper reports 2,357 SJTs. Each then receives human preferences nominally from 10 people drawn from a pool of 550, with neutral counted as half a vote and N/A filtered.

For 25 LLMs, the authors sample 20 responses per scenario at temperature 1.0. The prompt requires exactly one of the two actions, forbids neutrality or recommending both, and limits the response to two sentences; Gemini 3 Flash classifies each text as agree, oppose, or neither. They compare the frequency of the trait-oriented action with the human distribution through Trait-Positive Rate, absolute difference, majority-choice consistency, and directional alignment around 0.5. They report that models remain above 90% consistency when human preference is near 50/50; that alignment improves with consensus and capability, although some frontier models disagree in 15%-20% of non-unanimous high-consensus cases; and that smaller models are often near chance. They also show gaps between direct 1-7 ratings and SJT behavior, especially for impulsiveness.

The interpretation should remain bounded to behavior elicited by this instrument. Forcing a single action removes the mixed and neutral responses at issue, so consistency across 20 samples is not calibrated epistemic confidence and the reported overconfidence is partly prompt-induced. Ten votes per scenario provide a coarse, predominantly US/UK preference reference rather than a universal social norm. The 2,357 SJTs are nested within 161 source statements and the same raters and models contribute repeated measurements, yet the paper provides no hierarchical analysis, intervals, cluster bootstrap, or inferential tables. Judge validation covers only 100 responses reported all correct, without class balance, agreement, blinding, or error analysis. The self-report/SJT comparison shows weak cross-format prediction under this design, not internal traits or general invalidity of every self-report protocol.

Official artifacts released after the preprint improve transparency but support only partial reproduction. Kaggle v3 contains 2,262 scenarios, 95 fewer than the paper's 2,357, and omits source statements, individual judgments, validation records, original outputs, and 25-model results; it does exactly reproduce the 1,348 high-consensus Figure 4 denominators. Moreover, 109 aggregate scores do not advance in half-point increments despite documentation claiming 10 annotators and neutral=0.5. The official notebook defaults to one Gemini model, 500 rows, and 6 replications rather than 25 models, 1,348 rows, and 20 replications; it is invalid strict JSON because of a trailing comma, and the code treats model TPR exactly equal to 0.5 asymmetrically relative to the paper's strict formula. The defensible contribution is a useful, partially reproducible SJT benchmark showing prompt-, sample-, and format-specific gaps, not latent dispositions or universal human misalignment.

Español

El estudio convierte ítems de cuestionarios psicológicos sobre empatía, regulación emocional, asertividad e impulsividad en pruebas de juicio situacional para asistentes. Parte de 332 enunciados, reduce manualmente el conjunto a 260, usa Gemini 3 Pro para filtrar y reformular, y retiene 161. Gemini 3 genera 16 escenarios por enunciado, 2.576 candidatos, condicionados a una clase provisional AGREE, OPPOSE o AMBIGUOUS. Tres anotadores deben aceptar por unanimidad que cada escenario contiene un dilema, que las dos acciones se oponen y que la acción agree refleja el enunciado; tras excluir el 8%, el paper declara 2.357 SJT. Cada uno recibe preferencias humanas supuestamente de 10 personas de un pool de 550, con neutral como medio voto y N/A filtrado.

Para 25 LLM, los autores muestrean 20 respuestas por escenario a temperatura 1,0. El prompt obliga a recomendar exactamente una de las dos acciones, prohíbe la neutralidad o ambas posibilidades y limita la salida a dos frases; Gemini 3 Flash clasifica cada texto como agree, oppose o neither. Comparan la frecuencia de la acción asociada al rasgo con la distribución humana mediante Trait-Positive Rate, diferencia absoluta, consistencia de la opción mayoritaria y alineamiento direccional respecto a 0,5. Reportan que los modelos mantienen más del 90% de consistencia cuando la preferencia humana está cerca de 50/50; que la alineación mejora con consenso y capacidad, aunque algunos modelos frontera discrepan en 15%-20% de casos con consenso no unánime; y que los modelos pequeños suelen acercarse al azar. También muestran diferencias entre ratings directos 1-7 y conducta en SJT, especialmente en impulsividad.

La interpretación debe quedarse en la conducta producida bajo este instrumento. Obligar una sola acción elimina precisamente respuestas mixtas o neutrales, por lo que la consistencia entre 20 muestras no equivale a confianza epistémica calibrada y la sobreconfianza observada está parcialmente inducida por el prompt. Diez votos por escenario ofrecen una referencia gruesa y mayoritariamente estadounidense/británica, no una norma social universal. Los 2.357 SJT están anidados en 161 enunciados y los mismos raters y modelos generan medidas repetidas, pero no hay análisis jerárquico, intervalos, bootstrap por clúster ni tablas inferenciales. La validación del juez usa solo 100 respuestas declaradas todas correctas, sin balance de clases, acuerdo, cegamiento o análisis de errores. La comparación self-report/SJT muestra debilidad predictiva entre formatos en este diseño, no rasgos internos ni invalidez general de todo autoinforme.

Los artefactos oficiales publicados después del preprint mejoran la transparencia, pero solo permiten reproducción parcial. El CSV de Kaggle v3 contiene 2.262 escenarios, 95 menos que los 2.357 del paper, y omite enunciados fuente, juicios individuales, validaciones, salidas originales y resultados de los 25 modelos; sí reproduce exactamente los 1.348 denominadores de alto consenso de la Figura 4. Además, 109 scores no avanzan en medios puntos pese a que la documentación afirma 10 anotadores y neutral=0,5. El cuaderno oficial usa por defecto un Gemini, 500 filas y 6 repeticiones, no 25 modelos, 1.348 filas y 20 repeticiones; contiene JSON inválido por una coma final y el código trata asimétricamente un TPR de modelo igual a 0,5 frente a la fórmula estricta del paper. La aportación defendible es un benchmark SJT útil y parcialmente reproducible que revela brechas específicas de prompt, muestra y formato, no disposiciones latentes ni desalineamiento humano universal.

Research question

To what extent do the recommendations of 25 LLMs in situational judgment tests derived from questionnaires coincide with the distribution of human preferences, and do direct self-reports predict observed behavior in those scenarios?

Method

Experimental preprint with 161 retained items from four traits and 16 SJTs generated per item. After unanimous validation by three annotators, it declares 2,357 scenarios with 10 human preferences per SJT from a pool of 550. It evaluates 25 LLMs with 20 forced generations choosing one action, uses Gemini 3 Flash as judge, and computes Trait-Positive Rate, misalignment, confidence/consensus, and Directional Alignment. It also compares self-report ratings 1-7 with SJT behavior. Code and partial data were officially published in May 2026.

Sample: The paper declares 2,357 SJTs derived from 161 items, 10 preferences per scenario from a pool of 550 raters, and 20 responses per SJT for each of 25 models: 1,178,500 planned generations before failures and exclusions. The demographic table sums 22,883 ratings, 687 fewer than 2,357x10. The public release contains 2,262 SJTs, but retains exactly the 1,348 high-consensus cases and their cells from Figure 4.

Findings

  • The paper reports that all 25 models maintain more than 90% choice consistency when the human preference is near 50/50.
  • Directional alignment improves with human consensus and model capability; models smaller than 25B frequently appear near chance.
  • Even frontier models remain approximately in the low-mid 80s across several cells of 8/10 and 9/10 consensus, according to the figure.
  • Qualitative examples show emotional openness versus composure, harmony versus defending one's position, and immediacy versus logistical verification.
  • All models self-rate below 4/7 on impulsivity, while the majority exceeds 50% impulsive behavior in the selected SJTs.
  • Relative positions among models in self-report and SJT frequently do not coincide; some per-model graphs show negative trends.
  • The public file reproduces the 1,348 denominators by trait and consensus band from Figure 4.
  • The published magnitudes do not include intervals or numerical tables that allow evaluating uncertainty or small differences.

Limitations

  • The prompt forces one action, prohibits neutrality and recommending both, so consistency does not equate to calibrated confidence.
  • Model TPR is frequency across 20 samples at temperature 1.0, dependent on the sampler and the prompt.
  • Ten votes per SJT produce coarse estimates and their uncertainty is not propagated.
  • The human sample is concentrated in the United States and the United Kingdom and does not represent a universal norm.
  • No recruitment platform, payment, consent/IRB, eligibility, rater workload, assignment, or quality control is reported.
  • The scenarios are nested within 161 items and there are repeated measures of raters and models without hierarchical analysis or clustered bootstrap.
  • No intervals, tests, coefficients, effect sizes, or multiplicity correction are published for the significance claims.
  • The judge is validated with only 100 responses, without balance, stratification, agreement, blinding, or error interval.
  • Gemini participates in screening, generation, judgment, and the evaluated set, without analysis with alternative generators or judges.
  • The three-annotator validation checks coherence and mapping, not psychometric equivalence, reliability, factor structure, or invariance.
  • The provisional target class conditions context generation toward agree, oppose, or ambiguous.
  • The three qualitative cases are selected without a sampling protocol or coding.
  • Stable API IDs, serving providers, collection dates, system prompts, seeds, and complete inference details are missing.
  • The release has 2,262 SJTs versus 2,357 declared and omits 95 cases, source items, and individual judgments.
  • The demographic table sums 22,883 judgments, 687 fewer than the 23,570 expected; the missingness is not reconciled.
  • 109 public scores use increments incompatible with exactly 10 votes and neutral=0.5; there are no individual data to resolve it.
  • No generations, judge outputs, self-report ratings, or result matrices for the 25 models are released.
  • The default notebook evaluates one model, 500 rows, and 6 replicas, and only implements a Gemini runner.
  • The notebook is not strict JSON due to a trailing comma and requirements.txt does not pin versions.
  • The code considers TPR=0.5 aligned only on the agree side, in conflict with the strict inequality and with tie asymmetry.
  • The dataset license is CC BY-NC 4.0, different from CC BY 4.0 of the preprint and Apache-2.0 of the code.

What the study does not establish

  • It does not establish personality, mind, beliefs, or stable internal dispositions in the models.
  • It does not establish calibrated epistemic overconfidence outside the forced-choice prompt.
  • It does not establish that the majority preference of ten raters is the universally correct action.
  • It does not establish alignment with all of humanity or cross-cultural generalization.
  • It does not demonstrate that all self-report protocols lack predictive validity.
  • It does not isolate training or alignment as the cause of differences between models.
  • It does not allow precise rankings or significance of small differences without statistical uncertainty.
  • It does not validate the transformed SJTs as a psychometric scale equivalent to the source questionnaires.
  • It does not exactly reproduce the original results of 25 models with the current public artifacts.
  • It does not demonstrate behavior in open conversations where the model can express uncertainty, nuances, or alternatives.

Traceability

Scope: Full text

Version: arXiv:2602.11328v1, submitted and last updated 2026-02-11; CC BY 4.0

Consulted source: https://arxiv.org/abs/2602.11328

Review: Codex 26-page visual full-text, complete arXiv source, SJT construction, human-ground-truth, forced-choice, nested-design, judge, model-reporting, official code, Kaggle data and claim-boundary audit, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • Gemini 3 Pro
  • GPT-5.1
  • Claude 4 Sonnet
  • Grok 4 (0709)
  • Gemini 3 Flash
  • Claude 3.5 Haiku
  • Grok 3 Mini
  • DeepSeek R1
  • DeepSeek V3
  • Kimi K2
  • Mistral Large
  • Grok 4 Fast Reasoning
  • Grok Code Fast 1
  • Perplexity Sonar
  • GPT-OSS 120B
  • Codestral
  • GPT-OSS 20B
  • Gemma 3 12B
  • Qwen 3 8B
  • Ministral 8B
  • Gemma 3 4B
  • Ministral 3B
  • Qwen 3.1 7B
  • Gemma 3 1B
  • Qwen 3.0 0.6B

Instruments and metrics

  • Interpersonal Reactivity Index (IRI) para empatía
  • Emotion Regulation Questionnaire (ERQ)
  • Rathus Assertiveness Schedule (RAS)
  • Impulsive Behavior Short Scale-8 (I-8)
  • Trait Emotional Intelligence como marco de selección de cuatro rasgos
  • 2.357 Situational Judgment Tests declarados, derivados de 161 enunciados
  • Validación unánime de cada SJT por tres anotadores
  • Preferencia humana con agree, oppose, neutral y N/A
  • Prompt de acción abierta pero forced-choice, máximo dos frases
  • Gemini 3 Flash como LLM-as-a-Judge
  • Trait-Positive Rate, Trait Misalignment, confidence/consensus y Directional Alignment
  • Autoinforme 1-7 sobre los enunciados de preferencia

Data used

  • Conjunto final del paper: 2.357 SJT; no publicado íntegramente
  • Kaggle Behavioral Dispositions v3: 2.262 SJT, CC BY-NC 4.0
  • 23.000 anotaciones humanas reportadas de forma redondeada; juicios individuales no publicados
  • Matriz original de respuestas y clasificaciones de 25 modelos no publicada
  • Código Apache-2.0 behavioral_dispositions en google-research, commit 3fad2ebe5925da4d14c94f9e5178ad0d49eb4ccd

Evidence and location

  • Design, prompts, metrics, results, limitations, tables, and appendices: arXiv:2602.11328v1, all 26/26 PDF pages rendered and individually inspected
  • Version, date, authorship, category, and license: Official arXiv abstract and Atom metadata inspected 2026-07-18
  • Complete preprint materials and absence of hidden artifacts in v1: Complete official arXiv v1 source archive sha256 e6f781a73a7314477373526d5c458ce1972c9e65dd06b480ea742a37d5503d98
  • Subsequent publication, scope, and defects of reproducible code: Official google-research commit 3fad2ebe5925da4d14c94f9e5178ad0d49eb4ccd and all behavioral_dispositions source files inspected
  • Rows, schema, license, scores, and denominator match of Figure 4: Official Kaggle Behavioral Dispositions v3 CSV sha256 9e7edebdc925dfb07902113d7b1374a89637e09218bcca2b29f42608c5cfce90
  • Audit of sample, measurement, forced choice, nesting, judge, artifacts, code, and limits: reports/verification/article-401-sjt-ground-truth-forced-choice-confidence-nesting-judge-code-data-and-claim-audit.json