Decoding Emergent Big Five Traits in Large Language Models: Temperature-Dependent Expression and Architectural Clustering

Evaluation and psychometric validity2025ACL AnthologyApproved editorial review

Authors: Christos-Nikolaos Zacharopoulos, Revekka Kyriakoglou

Keywords: Large Language Models, Personality, Psychometrics, Model Evaluation, LLM Evaluation

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

2
Authors
13
Findings
19
Limitations
10
Evidence

Editorial summary

English

The reference version is the paper published in Findings of IJCNLP-AACL 2025. It administers the 60 BFI-2 items one prompt at a time, requesting a numerical 1–5 answer, to systems labeled Llama 3 8B, Mistral 7B, MythoMax L2 13B, Gemma 9B, Qwen 7B, and StripedHyena 7B. The paper reports 21 temperatures from 0 to 2 and one answer per item, model, and temperature: 7,560 cells with no replications. It incorrectly says increments of 1; code and CSVs confirm 0.1 steps. It reports Kruskal–Wallis differences for Extraversion, Agreeableness, Conscientiousness, and Openness but not Neuroticism, plus a negative temperature correlation with Neuroticism and a positive one with Extraversion. Its cluster places Gemma–Mistral together, Llama–StripedHyena in another branch, and Qwen at the opposite edge. The OSF audit, however, prevents a confirmatory interpretation. The supposed Qwen 7B is actually `Qwen/Qwen1.5-72B-Chat` in both code and all 1,260 CSV rows, so the sample is not six comparable 7B–13B models. The executable BFI-2 reverse-key list omits items 3, 4, 8, 9, 37, 42, 47, 48, and 58 and wrongly reverses item 46. Another script header prints the correct key, but calculations import the wrong one. Four domains and every associated mean, test, regression, figure, and cluster are affected; only Openness has the complete correct key. Released code and six CSVs also fail to reproduce the tables. The released key yields H=47.0473/85.1752/79.0852/13.7920/72.8568 rather than 40.7803/65.3067/63.0415/9.2691/58.1957. Released temperature correlations are r=-.4821 for Neuroticism and .5194 for Extraversion rather than -.5904 and .5021. Correct scoring makes Neuroticism H=44.8853, p=1.53e-8, contradicting the paper's central null. The aggregate Extraversion association falls to r=.4438; excluding the mislabeled 72B Qwen makes it r=.3757, p=.0933 and non-significant. There are also 265 NaN responses among 7,560 cells: 200 for StripedHyena and 51 for Qwen. StripedHyena loses as many as 36 of 60 items at one temperature. Pandas silently averages nonmissing items, producing unequal denominators without a documented missingness rule. The regression is not multiple linear regression by model: code first averages all six systems at each temperature and fits one line to only 21 ecological means. Kruskal–Wallis treats the 21 ordered conditions from the same questionnaire as independent samples for each model. One generation per cell cannot estimate sampling noise, test–retest reliability, or stability; changing temperature is not replication. The paper omits top-k=60, top-p=.8, repetition penalty=1.1, max tokens=20, and conversion of errors or invalid formats to NaN. It provides no factor structure, reliability, invariance, post-hoc comparison, or confidence intervals. A dendrogram of six profiles confounded by size, data, tuning, tokenizer, vocabulary, and attention cannot identify architectural predisposition. The final evidence is an open exploratory numeric-response dataset whose artifact exposes serious errors; it does not establish emergent or stable traits, temperature causality, architecture effects, or a governance basis.

Español

La versión de referencia es el artículo publicado en Findings of IJCNLP-AACL 2025. Administra los 60 ítems BFI-2, uno por prompt y con respuesta numérica de 1 a 5, a seis sistemas etiquetados como Llama 3 8B, Mistral 7B, MythoMax L2 13B, Gemma 9B, Qwen 7B y StripedHyena 7B. El paper reporta 21 temperaturas entre 0 y 2, una respuesta por ítem, modelo y temperatura: 7.560 celdas sin repeticiones. Escribe por error increments of 1; el código y los CSV confirman pasos de 0,1. Publica diferencias Kruskal–Wallis en Extraversion, Agreeableness, Conscientiousness y Openness, pero no Neuroticism, y correlaciones de temperatura negativas con Neuroticism y positivas con Extraversion. El clustering sitúa Gemma–Mistral juntos, Llama–StripedHyena en otra rama y Qwen en el extremo opuesto. Sin embargo, la auditoría del OSF invalida una lectura confirmatoria. El supuesto Qwen 7B es realmente `Qwen/Qwen1.5-72B-Chat`, tanto en el código como en las 1.260 filas del CSV; por tanto, la muestra no contiene seis modelos comparables de 7B–13B. La clave ejecutable de ítems invertidos del BFI-2 omite 3, 4, 8, 9, 37, 42, 47, 48 y 58 e invierte indebidamente el 46. El propio encabezado de otro script imprime la clave correcta, pero los cálculos importan la errónea. Esto contamina cuatro dominios, todas sus medias, pruebas, regresiones, figuras y clusters; solo Openness queda correctamente invertido. Además, el código y los seis CSV publicados no reproducen las tablas. Con la clave liberada se obtienen H=47,0473/85,1752/79,0852/13,7920/72,8568, no 40,7803/65,3067/63,0415/9,2691/58,1957. Las correlaciones liberadas son r=-0,4821 para Neuroticism y 0,5194 para Extraversion, no -0,5904 y 0,5021. Al corregir la clave, Neuroticism pasa a H=44,8853, p=1,53e-8, contradiciendo el principal resultado nulo del paper. La asociación agregada con Extraversion baja a r=0,4438; si se excluye el Qwen 72B mal rotulado, queda r=0,3757, p=0,0933 y deja de ser significativa. Hay además 265 respuestas NaN de 7.560: 200 en StripedHyena y 51 en Qwen. StripedHyena pierde hasta 36 de 60 ítems en una temperatura. Pandas calcula medias ignorando esos NaN, con denominadores variables y sin regla de missingness documentada. La regresión tampoco es multiple linear regression por modelo: el código promedia primero los seis sistemas en cada temperatura y ajusta una recta a solo 21 medias ecológicas. Kruskal–Wallis trata las 21 condiciones ordenadas del mismo cuestionario como muestras independientes de cada modelo. Una sola generación por celda no permite estimar aleatoriedad, test-retest o estabilidad; variar temperatura no sustituye repetir una condición. El paper no reporta top-k=60, top-p=0,8, repetition penalty=1,1, max tokens=20 ni que errores y formatos inválidos se convierten a NaN. Tampoco ofrece estructura factorial, fiabilidad, invariancia, comparación post-hoc o intervalos. Un dendrograma de seis perfiles confundidos por tamaño, datos, tuning, tokenizer, vocabulario y atención no identifica predisposición arquitectónica. La evidencia final es un conjunto abierto de respuestas numéricas exploratorias cuyo artefacto permite descubrir errores graves; no demuestra rasgos emergentes o estables, causalidad de temperatura, efectos de arquitectura ni una base para gobernanza.

Research question

Do six systems labeled as 7B-13B LLMs differ in BFI-2 responses, does their aggregated mean change with temperature, and do they form descriptively clusterable profiles?

Method

60 BFI-2 responses are generated per system across 21 temperatures from 0.0 to 2.0, once per cell, using Together with top-k=60, top-p=0.8, repetition penalty=1.1, and max tokens=20. The artifact computes domains with an erroneous inverted key, applies Kruskal-Wallis to 21 conditions per label, averages the six models per temperature for a 21-point regression, and clusters six mean profiles with Ward/euclidean distance.

Sample: Six labels, 60 items, and 21 temperatures 0.0-2.0 in actual steps of 0.1: 7,560 cells and a single generation per combination. There are 7,295 valid responses and 265 NaN: Gemma 6, MythoMax 8, Qwen 72B mislabeled 51, and StripedHyena 200. There are no human participants, annotators, repeated seeds, or test-retest.

Findings

  • The current source is Findings of IJCNLP-AACL 2025, Anthology ID 2025.findings-ijcnlp.104, DOI 10.18653/v1/2025.findings-ijcnlp.104, pages 1678-1685.
  • The 8 published pages were rendered and visually inspected; SHA-256 d1662c1ed14904e8e06421c601746911fd46790e2fc4ba52c6ae10ad78f060a1.
  • Extraversion, Agreeableness, Conscientiousness, and Openness differ globally across models according to Kruskal-Wallis (H = 40.7803; 65.3067; 63.0415; 58.1957; p < 0.01); Neuroticism does not (H = 9.2691).
  • Temperature is negatively associated with Neuroticism (R² = 0.3486, r = -0.5904, p < 0.05) and positively with Extraversion (R² = 0.2521, r = 0.5021, p < 0.05).
  • Agreeableness (R² = 0.0343), Conscientiousness (0.0257), and Openness (0.0003) show no significant association with temperature in the aggregated analysis.
  • Clustering first joins Gemma-Mistral, adds MythoMax, separates Qwen, and forms another Llama-StripedHyena pair; it is a descriptive structure of only six model observations.
  • The code and CSVs confirm temperature steps of 0.1 and resolve the increments of 1 error in the published method.
  • The system called Qwen 7B is Qwen/Qwen1.5-72B-Chat in the code and the 1,260 rows of the CSV.
  • The executed key omits nine inverted BFI-2 items and improperly inverts item 46.
  • The artifact does not reproduce any H from Table 1 or the central correlations/R² from Table 2.
  • With correct scoring, Neuroticism gives H=44.8853, p=1.53e-8, not the published null result.
  • With correct scoring and without the mislabeled Qwen 72B, Extraversion-temperature yields r=0.3757, p=0.0933.
  • The regression published in code uses 21 means across models, not separate slopes or a multilevel model.

Limitations

  • The paper writes increments of 1; only the artifact reveals they were steps of 0.1.
  • Qwen 7B is actually Qwen1.5-72B-Chat, invalidating the description of comparable sizes.
  • The executed BFI-2 key has nine omissions and one extra inversion.
  • The published tables are not reproduced with the linked code and CSVs.
  • Results change materially when correcting scoring and excluding the erroneous Qwen label.
  • There are 265 NaN with strong concentration in StripedHyena and Qwen.
  • Means ignore NaN and use variable denominators without a missingness analysis.
  • The current code has the Llama entries commented out and does not regenerate the six CSVs from start to finish.
  • Together, torch, and service snapshots are not pinned in a reproducible environment.
  • The paper omits top-k, top-p, repetition penalty, max tokens, and error handling as NaN.
  • A single response per item and temperature does not allow estimating test-retest or separating sampling error from a systematic change, especially at high temperatures.
  • The regression averages the six models per temperature and fits only 21 points; it does not estimate intercepts, slopes, or interactions per model.
  • Kruskal-Wallis treats 21 ordered conditions of the same questionnaire as independent observations per label.
  • Five Kruskal-Wallis tests lack multiple correction, post-hoc, and effect sizes; a global p does not identify which pairs differ or their practical relevance.
  • The unpublished pairwise code calls rank-biserial U/(n1×n2), which is not that transformation, and does not return Bonferroni-adjusted p values.
  • BFI-2 is validated for people, not for tokens; reliability, factor structure, convergence, or invariance is not tested.
  • Architecture, data, tuning, size, vocabulary, and tokenizer are confounded. A dendrogram of six models cannot causally attribute profiles to attention mechanisms.
  • Clustering has no bootstrap, scale/distance sensitivity, uncertainty, or external validation.
  • The study is limited to one prompt, English, and numeric responses; it does not cover dialogue, behavior, or other languages.

What the study does not establish

  • It does not demonstrate that models possess traits, emotions, creativity, internal personality, or a human psychological baseline.
  • It does not demonstrate that temperature causes Extraversion or Neuroticism; it identifies score associations in an exploratory and aggregated design.
  • It does not establish personality stability, because there are no independent repetitions or test-retest.
  • It does not allow attributing clusters to architecture or training data, or separating nature from learning.
  • It does not validate BFI-2 for comparing LLMs, nor does it make their means equivalent to human scores.
  • It does not demonstrate the null result for Neuroticism with correct scoring.
  • It does not demonstrate a robust temperature-Extraversion association without the mislabeled Qwen 72B.
  • It does not reproduce the published tables from the official artifact.
  • It does not test that these profiles improve selection, tuning, safety, or governance in real applications.

Traceability

Scope: Full text

Version: Findings of IJCNLP-AACL 2025, Anthology ID 2025.findings-ijcnlp.104, DOI 10.18653/v1/2025.findings-ijcnlp.104, pages 1678-1685, 8 pages; supersedes arXiv:2511.04499v1

Consulted source: https://aclanthology.org/2025.findings-ijcnlp.104/

Review: Codex complete bilingual full-text fidelity pass using the published Findings version, all-page visual inspection, OSF archive inventory by HTTP range, source-code audit, six-CSV missingness audit, exact reproduction of released analysis, corrected BFI-2 rescoring, mislabeled-model sensitivity analysis, statistical-design review, and metadata reconciliation; summaries written from the paper and executable evidence rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • meta-llama/Meta-Llama-3-8B-Instruct-Lite in released CSV; collection entry commented out in current code
  • mistralai/Mistral-7B-Instruct-v0.3
  • Gryphe/MythoMax-L2-13b-Lite
  • google/gemma-2-9b-it
  • Qwen/Qwen1.5-72B-Chat, incorrectly labeled Qwen 7B in paper, code display name, path, and figures
  • togethercomputer/StripedHyena-Nous-7B

Instruments and metrics

  • Big Five Inventory-2, 60 items, with materially incorrect executable reverse-scoring key
  • One-item-at-a-time numeric 1-5 prompt
  • Kruskal–Wallis H tests across 21 temperature scores per model label
  • Pearson correlation and simple regression on 21 model-averaged temperature means
  • Ward agglomerative hierarchical clustering with Euclidean distance over six five-domain profiles
  • Independent artifact reanalysis with released and corrected BFI-2 scoring keys

Data used

  • OSF bsvzc Data and Code for LLM Personality: six analyzed CSVs, 7,560 cells, 265 NaN responses
  • OSF source code, BFI-2 questionnaire, figures, environment file, additional unused model CSVs, and bundled model archive
  • Published Findings of IJCNLP-AACL 2025 tables that do not reproduce from the released code and six analyzed CSVs

Evidence and location

  • Published version, DOI, pages, and abstract: ACL Anthology record and Findings of IJCNLP-AACL 2025 page 1678 checked 15 July 2026
  • Design, model labels, increment error, and declared analysis: Published pages 1678-1679, Introduction and Methods
  • Published Kruskal-Wallis, regression, and clustering results: Published pages 1679-1681, Tables 1-2 and Figures 1-2
  • Associative and non-causal caveats acknowledged: Published pages 1680-1682, Discussion and Limitations
  • Exact prompt: Published page 1684, Appendix A
  • Descriptive means per label: Published page 1685, Appendix Figure 3
  • Temperatures 0.1, BFI-2 keys, Together parameters, aggregation, and clustering: OSF bsvzc src/config.py, src/llm_prompting.py and src/big_five_analysis.py inspected 15 July 2026
  • Mislabeled Qwen 72B, 7,560 cells, and 265 NaN: Six OSF result CSVs inspected and recomputed 15 July 2026
  • Non-reproducible results and reanalysis with correct key: reports/verification/article-190-scoring-and-artifact-reproduction-audit.json
  • Complete visual inspection: All 8 published PDF pages rendered and visually inspected on 15 July 2026