Persistent Instability in LLM's Personality Measurements: Effects of Scale, Reasoning, and Conversation History

Evaluation and psychometric validity2026AAAI ProceedingsApproved editorial review

Authors: Tommaso Tosato, Saskia Helbling, Yorguin-Jose Mantilla-Ramos, Mahmood Hegazy, Alberto Tosato, David John Lemay, Irina Rish, Guillaume Dumas

Keywords: Computation and Language, Artificial Intelligence, Large Language Models, Personality Measurements, Model Behavior Consistency

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

8
Authors
16
Findings
28
Limitations
17
Evidence

Editorial summary

English

PERSIST studies how much a model's response to the same personality item changes when the administration context changes. It combines the 44-item BFI and 27-item Short Dark Triad, 71 items and eight traits in total, with LLM-adapted versions, random order, paraphrases, five highlighted persona profiles, reasoning on or off, and conversation history. The method enumerates 29 models from eight families, spanning 1B to 671B parameters, although the abstract and introduction claim 25 open-source models spanning 1B to 685B; the paper does not reconcile these counts. Main experiments use temperature 0, a reported 250 runs, and one question per turn. Reasoning experiments use temperature 0.6, and Claude is limited to 70 runs. The central metric is the standard deviation of each 1–5 item response across runs, later aggregated by model and trait.

The descriptive pattern is clear. Question-level variability decreases with model size (Spearman, p<0.001), but does not vanish, and the authors display standard deviations above 0.3 even for models larger than 400B. Perplexity does not change with size (p=0.934), although it correlates moderately with question-level SD (rho=0.465). Mean scores become higher for Openness, Conscientiousness, Extraversion, and Agreeableness and lower for Neuroticism and Dark Triad traits as models scale. This can reflect alignment, extremity, and ceiling effects rather than a more prosocial personality. LLM-adapted instruments do not reduce SD relative to the originals (Wilcoxon p=0.286) and increase perplexity (p<0.001), so removing human-specific references is not sufficient to stabilize responses.

Interventions have heterogeneous effects. Greater reasoning effort raises SD in GPT-OSS; Qwen 3, Qwen 3 MoE, DeepSeek, and Claude show higher variability with reasoning in the aggregated analyses, while perplexity usually falls. Buddhist and teacher personas reduce SD relative to the assistant; the schizophrenia persona increases both SD and perplexity, but the antisocial persona does not significantly change SD (p=0.260), so the abstract overstates the result when it generalizes higher variability to plural 'misaligned personas.' Paraphrases increase variability only in the published group of four models at least 50B (p<0.01; smaller models p=0.244). Conversation history raises SD for the 19 models below 50B and lowers it for the four larger models; the change correlates negatively with size (rho=-0.512, p=0.0126).

The artifact exposes questionnaires, prompts, generation code, and analysis scripts, but it does not reproduce the results: the official repository contains no raw outputs, preprocessed CSVs, figures, or logs, and has no data release. The current code also treats model-by-question differences as independent observations in the paraphrase and history Wilcoxon tests even though Tables 3 and 4 label n as the number of models; the reasoning analysis likewise applies Mann–Whitney to question-level values although many comparisons are paired. This pseudoreplication can make p-values too optimistic. The generator interprets n_iter inclusively: n_iter=250 creates 251 conditions and n_iter=100 creates 101, while no run configurations or data are available to establish what was actually executed. The BFI bank contains 4,399 paraphrases, 4,377 unique strings, and only 99 entries for 'Can be moody,' not 100.

The evidence supports the conclusion that these 1–5 self-reports are sensitive to context, wording, and inference mode, and that a single score is not a stable measurement. It does not establish that models lack every form of behavioral consistency, that the selected SD predicts real safety failures, or that an architectural deficiency has been identified. Neither the original nor adapted questionnaires has psychometric validation for LLMs, no external behavior is tested, and SD >0.3 is not justified as a hazardous threshold. PERSIST is a useful stress test for administration fragility; its safety conclusions should be read as motivated hypotheses rather than empirical risk certification.

Español

PERSIST estudia cuánto cambia la respuesta de un modelo al mismo ítem de personalidad cuando varía el contexto de administración. Combina BFI-44 y Short Dark Triad, 71 ítems y ocho rasgos en total, con versiones reformuladas para LLM, orden aleatorio, paráfrasis, cinco perfiles de persona destacados, razonamiento activado o desactivado e historial conversacional. El método enumera 29 modelos de ocho familias, de 1B a 671B parámetros, aunque el abstract y la introducción dicen 25 modelos open-source de 1B a 685B; el paper no reconcilia esas cifras. Las ejecuciones principales usan temperatura 0, 250 runs declarados y preguntas de una en una; los experimentos de razonamiento usan temperatura 0,6 y Claude se limita a 70 runs. La métrica central es la desviación estándar de la respuesta 1–5 a cada ítem entre ejecuciones, agregada después por modelo y rasgo.

El patrón descriptivo es claro. La variabilidad por pregunta disminuye al aumentar el tamaño del modelo (Spearman, p<0,001), pero no desaparece y los autores muestran SD superiores a 0,3 incluso en modelos de más de 400B. La perplejidad no cambia con el tamaño (p=0,934), aunque correlaciona moderadamente con la SD por pregunta (rho=0,465). Los scores medios se vuelven más altos en apertura, consciencia, extraversión y amabilidad y más bajos en neuroticismo y Dark Triad al escalar; esto puede reflejar alignment, extremidad y efectos techo, no una personalidad más prosocial. Las versiones adaptadas a LLM no reducen la SD frente a los cuestionarios originales (Wilcoxon p=0,286) y elevan la perplejidad (p<0,001), por lo que eliminar referencias humanas no basta para estabilizar respuestas.

Las intervenciones producen efectos heterogéneos. En GPT-OSS, más reasoning effort eleva la SD; Qwen 3, Qwen 3 MoE, DeepSeek y Claude muestran mayor variabilidad con razonamiento en los análisis agregados, mientras la perplejidad suele bajar. Las personas buddhist y teacher reducen la SD frente al assistant; schizophrenia la aumenta junto con la perplejidad, pero antisocial no cambia significativamente la SD (p=0,260), de modo que el abstract exagera al generalizar el aumento a «misaligned personas» en plural. Las paráfrasis aumentan la variabilidad solo en el grupo de cuatro modelos de al menos 50B según el contraste publicado (p<0,01; modelos menores p=0,244). El historial conversacional aumenta la SD en los 19 modelos menores de 50B y la reduce en los cuatro mayores; el delta se correlaciona negativamente con tamaño (rho=-0,512, p=0,0126).

El artefacto permite inspeccionar cuestionarios, prompts, generador y scripts, pero no reproducir los resultados: el repositorio oficial no contiene outputs brutos, CSV preprocesados, figuras ni logs, y no ofrece release de datos. Además, el código actual trata las diferencias modelo×pregunta como observaciones independientes en los Wilcoxon de paráfrasis e historial, aunque las Tablas 3–4 presentan n como número de modelos; el análisis de razonamiento también usa Mann–Whitney sobre preguntas pese a que muchas comparaciones están emparejadas. Esa pseudorreplicación puede hacer los p-valores demasiado optimistas. El generador interpreta n_iter de forma inclusiva: n_iter=250 produce 251 condiciones y n_iter=100 produce 101, sin que existan configuraciones de ejecución para saber qué se lanzó realmente. El banco BFI contiene 4.399 paráfrasis, 4.377 únicas, y «Can be moody» solo tiene 99, no 100.

La evidencia respalda que estos autoinformes 1–5 son sensibles al contexto, al wording y al modo de inferencia, y que un score único no representa una medida estable. No demuestra que los modelos carezcan de toda consistencia conductual, que la SD elegida prediga fallos de seguridad reales o que exista una carencia arquitectónica. Los cuestionarios originales y adaptados carecen de validación psicométrica para LLM, no se evalúa conducta externa y el umbral SD>0,3 no se justifica como peligroso. PERSIST es una batería útil para detectar fragilidad de administración; sus conclusiones de seguridad deben leerse como hipótesis motivadas, no como certificación empírica de riesgo.

Research question

To what extent are the BFI and Dark Triad scores of LLMs stable under changes of size, persona, reasoning, wording, order and conversational history, and do scaling or adapted instruments reduce that instability?

Method

Multifactorial experimental battery. BFI-44 and SD3-27, together with BFI-LLM and SD3-LLM, are administered to a declared list of 29 models. Questions are answered individually on a 1–5 scale under 250 declared runs of order, 100 declared paraphrases, personas assistant/buddhist/teacher/antisocial/schizophrenia, and depression in supplementary figures, with and without history and with several reasoning modes. Mean per trait, SD of each item across runs and perplexity of the response token are calculated. Spearman, Wilcoxon, Kruskal–Wallis/Dunn and Mann–Whitney correlations are applied. The audit read the official version and the extended version, rendered tables and figures, reviewed the official repository at cee3b5ff8291bd794b0f751f10f9f2761fdccfd7, statically compiled the Python, counted questionnaires/paraphrases and ran the generator without inference to check its iteration semantics.

Sample: There are no new human participants. Each model response to an item is an observation, repeated under permutations or paraphrases. The method enumerates 29 models; the main scaling figures show 24 configurations and the reasoning experiments add specific families/modes. All experiments are declared with 250 runs except Claude, with 70, but the public generator creates n_iter+1 conditions. Each run with an invalid response is excluded entirely: the extended version reports 0% invalid runs for scaling, paraphrase and history, 0.07% for personas and 0.35% for reasoning.

Findings

  • The mean SD per question decreases with model size (Spearman p<0.001), while perplexity shows no relationship with size (p=0.934).
  • Perplexity and per-question SD correlate moderately in the scale analysis (rho=0.465), so token uncertainty explains only part of the variation.
  • Even models larger than 400B show SDs above 0.3 on the 1–5 scale in the figures; the work does not define an external threshold of acceptability.
  • With size, positive scores rise (p=0.001) and negative scores fall (p<0.001), but the authors themselves warn of possible ceiling effects and departure from human norms.
  • The LLM-adapted versions do not significantly change SD relative to the originals (p=0.286) and increase perplexity (p<0.001).
  • GPT-OSS shows higher SD with greater reasoning effort; the aggregated analyses report more SD with reasoning in Qwen 3, Qwen 3 MoE, DeepSeek and Claude.
  • Perplexity usually drops when reasoning is activated even when SD rises, so greater local confidence does not imply greater stability across runs.
  • Buddhist reduces negative scores (p<0.001), SD (p=0.005) and perplexity (p=0.020) relative to assistant; the difference in positive scores is not significant.
  • Teacher increases positive scores, reduces negative scores and reduces SD (p=0.021), with no significant change in perplexity.
  • Antisocial shifts mean scores in the induced direction, but does not significantly change SD (p=0.260) or perplexity (p=0.098).
  • Schizophrenia reduces positive scores, raises negative scores, SD (p<0.05) and perplexity (p<0.001).
  • The effect of paraphrase on SD tends to grow with size (rho=0.39, p=0.0671); the published contrast is non-significant below 50B (n=19, p=0.244) and positive from 50B (n=4, p<0.01).
  • Conversational history has an SD delta inversely related to size (rho=-0.512, p=0.0126).
  • According to Tables 4, history raises SD below 50B and reduces it from 50B, both p<0.001; these p-values come from model×question differences, not from 19 and 4 independent model units.
  • The paper declares more than two million responses, but the repository does not publish the data needed to verify the total or recompute figures and tests.
  • The reported invalid run rate is low: 0.35% in reasoning, 0.07% in persona and 0% in the other three blocks.

Limitations

  • The abstract says 25 open-source models and 1B–685B, while the models section enumerates 29 including Claude and says 1B–671B; there is no inclusion table reconciling figures and experiments.
  • The section states that DeepSeek only participates in reasoning, but the scaling figures include DeepSeek V3.
  • The methodology enumerates five personas, while the supplement also plots depression; the exact scope of each block is not summarized in an experimental matrix.
  • No raw outputs, preprocessed data, generated figures, logs, exact commands, per-job seeds or model manifest are released; the results and the two-million total are not reproducible from the artifact.
  • The repository has only one public commit of initial release and no data release; README describes results folders that are ignored and absent.
  • The README examples call fig_scale.py, fig_persona.py, fig_reason.py and fig_paraph.py, files that do not exist with those names.
  • There are no automated tests, lockfile or weights/tokenizers reviews; pyproject uses broad ranges and omits scikit-posthocs, so Dunn is skipped if not installed.
  • The Llama 3.1 405B path is hardcoded to a private cluster scratch path, preventing that model from running without modifying code.
  • The generator uses an inclusive range: n_iter=250 creates 251 runs and n_iter=100 creates 101. Without configs/data it cannot be known whether the study compensated for the off-by-one.
  • Although 100 paraphrases per question are claimed, the BFI bank contains 99 for «Can be moody»; 15 BFI items and 3 SD3 items contain exact duplicates.
  • Paraphrases are generated with Qwen and reviewed by two authors, but no independent semantic validation, agreement, intensity, polarity or psychometric conservation per item is reported.
  • The BFI-LLM and SD3-LLM versions are face-valid adaptations by the authors with no factor analysis, internal consistency, invariance, convergent, discriminant or criterion validity.
  • The study measures self-report stability on an ordinal scale, not open behavior or decisions in healthcare, education, finance or medicine.
  • Question order only produces variation when history is preserved: order changes the accumulated context of questions and answers, so «reordering alone» does not describe an independent administration of items.
  • At temperature 0, order variation depends on conversational context and potential backend non-determinism; at temperature 0.6, reasoning mixes mode effect with stochastic sampling.
  • SD over categories 1–5 is descriptive, but has no validated risk threshold; >0.3 is not linked to operational error, harm or non-compliance of a decision.
  • The Wilcoxon tests for paraphrase and history group deltas of each question within models; these observations are nested and correlated, they are not independent replicates even though the tables highlight n of models.
  • The aggregated Mann–Whitney tests for reasoning compare per-question values as independent samples and do not exploit item/model pairing; DeepSeek R1 versus V3 additionally adds a model change.
  • No hierarchical model, model/family bootstrap, cluster permutation or leave-one-family-out is used that respects the dependence among questions, models and families.
  • Many p-values are reported with corrections not described coherently across tables and scripts, and almost never numerical effect sizes or intervals for the main contrasts.
  • The arbitrary 50B cutoff leaves only four large models; a significance obtained with many questions does not prove a general effect across model populations.
  • The correlation with size mixes families, dense and MoE and counts total parameters, for example 235B-A22B, without controlling for architecture, data, alignment or active parameters.
  • Human averages and their SDs describe variation between persons, while LLM bars describe variation between runs; they are not the same source of uncertainty nor a test of equivalence.
  • Human norms come from heterogeneous external samples and instruments and are used visually without demographic matching, inference or context correction.
  • Discarding an entire run for an invalid response avoids imputation, but may introduce selection conditioned on model/persona; the low aggregate rates do not show where discards concentrate.
  • Perplexity is calculated only on the parsed score token and is not available for Claude; it does not represent the complete uncertainty of reasoning or behavior.
  • The conclusion on architectural foundations is speculative: architecture, training or alignment is not intervened on, nor is there a comparison with a class designed for stability.
  • The paper acknowledges that strategic deception cannot be ruled out and that both instruments lack formal validation for LLMs.

What the study does not establish

  • It does not demonstrate that the 29 enumerated models participated in all experiments, nor does it clarify the exact set behind «25 open-source».
  • It does not independently verify the total of more than two million responses.
  • It does not demonstrate an internal personality, a stable self or their absence; it measures responses to inventories under specific prompts.
  • It does not demonstrate that all model behavior is unstable, nor that score instability predicts inconsistency in real tasks.
  • It does not demonstrate that SD>0.3 is dangerous, clinically relevant or incompatible with a use case.
  • It does not demonstrate that scaling can never improve stability; it observes limited gains in the included families and sizes.
  • It does not demonstrate that reasoning causes instability in general; the contrasts mix pairing, models and sampling at temperature 0.6.
  • It does not demonstrate that all misaligned personas increase variability: antisocial does not do so significantly in its own results.
  • It does not demonstrate that the LLM adaptation is valid; it only shows that those two reformulations do not reduce SD.
  • It does not completely separate the order effect from the accumulated history/context effect.
  • It does not prove that differences from human norms represent comparable personality differences.
  • It does not identify a causal neural or architectural mechanism, nor does it prove that current alignment strategies are globally inadequate.
  • It provides no evidence of safety incidents, harm or performance in high-risk applications.
  • It does not allow reproducing the published p-values without the missing data.
  • It does not replace a complete psychometric validation or a longitudinal behavioral evaluation.

Traceability

Scope: Full text

Version: arXiv:2508.04826v3 extended 18-page version, 23 Dec 2025; cross-checked against the official AAAI-26 proceedings paper, vol. 40(44), pp. 37961–37969, DOI 10.1609/aaai.v40i44.41133

Consulted source: https://arxiv.org/pdf/2508.04826v3

Review: Codex editorial review and methods/artifact audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Llama 3.1 Instruct: 8B, 70B, 405B
  • Qwen 2.5 Instruct: 1.5B, 3B, 7B, 14B, 32B, 72B
  • Qwen 3: 1.7B, 4B, 8B, 14B, 32B
  • Qwen 3 MoE: 30B-A3B, 235B-A22B
  • Gemma 2 Instruct: 2B, 9B, 27B
  • Gemma 3 Instruct: 1B, 4B, 12B, 27B
  • DeepSeek V3 and DeepSeek R1, reported as 671B
  • GPT-OSS 20B and 120B
  • Claude Sonnet 4.5 and Opus 4.1

Instruments and metrics

  • Big Five Inventory (BFI-44)
  • Short Dark Triad (SD3-27)
  • BFI-LLM, an author-adapted 44-item version
  • SD3-LLM, an author-adapted 27-item version
  • Five-point Likert responses from Strongly Disagree to Strongly Agree
  • Question-level standard deviation across runs
  • Trait-level mean score across items and runs
  • Token-response perplexity from extracted log probability
  • Spearman scaling and SD–perplexity correlations
  • Wilcoxon, Kruskal–Wallis with Dunn post hoc, and Mann–Whitney comparisons

Data used

  • 71 original questionnaire items: 44 BFI and 27 SD3
  • 71 author-written LLM-adapted items
  • BFI paraphrase bank: 4,399 entries and 4,377 unique strings in the released artifact
  • SD3 paraphrase bank: 2,700 entries and 2,697 unique strings in the released artifact
  • Reported 250 order runs and 100 paraphrasing settings
  • Reported total of more than 2 million individual responses
  • Human BFI and SD3 reference means from external norms
  • No raw model outputs or preprocessed analysis tables released

Evidence and location

  • Official publication, DOI, authors and declared scope: AAAI-26 official proceedings paper, vol. 40(44), pp. 37961–37969, DOI 10.1609/aaai.v40i44.41133; title page and abstract, p. 37961
  • 25 models, more than two million and five declared conclusions: Official paper abstract and introduction, pp. 37961–37962; arXiv v3 extended version, pp. 1–2
  • Framework, run discarding and processing: Official paper Methodology, p. 37963; extended version, p. 3
  • BFI, SD3 and adapted versions: Official paper Questionnaire Design, p. 37963; extended Appendix C, pp. 11–12
  • Factors, 29 models, runs, temperature and hardware: Official paper Experimental Design through Implementation Details, pp. 37963–37964; extended version, pp. 3–4
  • Scale, scores, SD, perplexity and rho 0.465: Official paper Effects of Model Scale and Table 1, p. 37964; extended version, pp. 4–5
  • Personas and exact contrasts: Official paper Effects of Persona Prompt, p. 37964; extended Tables A2–A5, p. 18
  • Reasoning and perplexity: Official paper Effects of Reasoning and Figure 2, pp. 37964–37966; extended version, pp. 4–6 and 14
  • Original instruments versus adapted ones: Official paper Table 2 and Figure 3, pp. 37964–37967; extended version, pp. 4, 7 and 15
  • Paraphrases and 50B cutoff: Official paper Table 3 and Figure 4, pp. 37964–37967; extended version, pp. 4, 7 and 16
  • Conversational history, rho and groups: Official paper Table 4 and Figure 5, pp. 37966–37967; extended version, pp. 6–7 and 17
  • Acknowledged limitations and conclusion: Official paper Limitations and Conclusion, p. 37968; extended version, p. 8
  • Prompts and invalid responses: arXiv v3 extended Appendix A–B and Table A1, pp. 10–11
  • Code, questionnaires, paraphrases and absence of data: Official tosato/PERSIST repository commit cee3b5ff8291bd794b0f751f10f9f2761fdccfd7, audited 15 Jul 2026; 31 tracked files, no raw or preprocessed results, no data release
  • n_iter semantics and paraphrase bank counts: Repository generator/utils.py plus direct non-inference execution at commit cee3b5f: n_iter 250 generated 251 response columns and n_iter 100 generated 101; JSON audit found BFI 4,399/4,377 total/unique and SD3 2,700/2,697
  • Unmodeled statistical dependence: Repository stats_convhist_trends.py and stats_paraph_comparison.py pool model×question std_diff in Wilcoxon; stats_reasoning_efforts.py uses aggregated question-level Mann–Whitney, audited 15 Jul 2026
  • Environment reproduction problems: Repository README, pyproject.toml and generator/generate.py: missing named figure scripts, no tests/lock/data, optional unpinned Dunn dependency and hardcoded 405B cluster path, audited 15 Jul 2026