Beyond Self-Reports: Multi-Observer Agents for Personality Assessment in Large Language Models

Evaluation and psychometric validity2025ACL AnthologyApproved editorial review

Authors: Yin Jou Huang, Rafik Hadfi

Keywords: Computation and Language, Artificial Intelligence, Personality Assessment, Multi-Observer Framework

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

2
Authors
12
Findings
23
Limitations
13
Evidence

Editorial summary

English

The paper adapts the psychological logic of informant reports to LLM agents. Instead of asking only a subject model to complete a questionnaire about itself, multiple observer agents converse with the subject in friendship, family, or workplace scenarios and then rate its behavior using 50 IPIP Big Five items. The main experiment creates 100 prompt-conditioned subjects, 15 observers per subject, five for each relational context, and five scenarios per subject–observer pair. GPT-4o generates relationships and scenarios and fills both subject and observer roles; the appendix repeats analyses with Qwen2.5-72B-Instruct and Llama-3-70B-Instruct. Ratings from groups of 5 to 15 observers are averaged and compared with the injected profile, self-report, and a small human evaluation.

The central result is a separation between instruction and simulated behavior. Self-report correlates almost perfectly with the trait profile inserted into the prompt, 0.93 to 0.97 depending on the dimension, whereas aggregated observer reports correlate less with that instruction, 0.55 to 0.86, but generally approximate human judgments of the dialogues more closely. In the main table, observer reports outperform self-report against human judgment for openness, agreeableness, and neuroticism and are slightly lower for conscientiousness and extraversion; in the appendix, their absolute distance is smaller for all five dimensions and all three models. Correlation curves stabilize at roughly five to seven observers. Systematic deviations also appear: observers rate agreeableness 0.91 points and conscientiousness 0.39 points above the subject’s self-rating, with Cohen’s d values of 1.07 and 0.46; family, friendship, and workplace contexts affect these two dimensions.

The valuable contribution is not evidence of internal personality but evidence that a declared profile can diverge from generated behavior. The qualitative case is illustrative: a subject instructed to be highly disagreeable gives itself an agreeableness score of 1.7, but produces a cautious and fairly cooperative dialogue and receives observer ratings around 3. The design supplements self-report with behavioral material, tests three model families, four prompt variants, reversed scale order and batch presentation, and adds a human comparison with inter-rater agreement. It is a promising methodological direction for measuring behavioral fidelity in induced personas.

However, the “informants” are not independent external observers. In the main condition, they are instances of the same GPT-4o that generates profiles, relationships, scenarios, both sides of the dialogue, and IPIP ratings. Observers and humans see the dialogues, while self-report directly receives the personality markers; this asymmetry structurally favors observer–human agreement and self-report–prompt agreement. The injected profile is not psychological ground truth, averaging correlated agents does not establish reliability, and the five-to-seven plateau is not validated with confidence intervals, resampling, or an external test set. Human evidence covers 16 cases with two raters per case, while the Responsible NLP checklist instead mentions 12 participants and confirms that no ethics-board approval or exemption was obtained. The repository declared available in the final paper returns 404, so prompts, outputs, and analyses are not currently reproducible. The defensible conclusion is that multi-agent reports better capture human impressions of these synthetic conversations under this protocol, not that they measure a true or more objective LLM personality.

Español

El trabajo adapta la lógica psicológica de los informant reports a agentes LLM: en vez de pedir solo al modelo sujeto que complete un cuestionario sobre sí mismo, hace que varios agentes observadores conversen con él en escenarios de amistad, familia o trabajo y después puntúen su conducta con 50 ítems IPIP del Big Five. El experimento principal crea 100 sujetos condicionados mediante perfiles, 15 observadores por sujeto, cinco por contexto relacional, y cinco escenarios por pareja. GPT-4o genera relaciones y escenarios y ocupa los roles de sujeto y observador; el apéndice repite análisis con Qwen2.5-72B-Instruct y Llama-3-70B-Instruct. Las puntuaciones de 5 a 15 observadores se promedian y se comparan con el perfil inyectado, el self-report y una pequeña evaluación humana.

El resultado central es una separación entre instrucción y conducta simulada. El self-report correlaciona casi perfectamente con el perfil de rasgos insertado en el prompt, 0,93 a 0,97 según dimensión, mientras el informe agregado de observadores correlaciona menos con esa instrucción, 0,55 a 0,86, pero suele aproximarse mejor a cómo humanos juzgan los diálogos. En la tabla principal, el observador supera al self-report frente al juicio humano en apertura, agradabilidad y neuroticismo, y queda ligeramente por debajo en consciencia y extraversión; en el apéndice, su distancia absoluta es menor para las cinco dimensiones y los tres modelos. Las curvas de correlación se estabilizan aproximadamente con cinco a siete observadores. También aparecen desviaciones sistemáticas: los observadores puntúan agradabilidad 0,91 puntos y consciencia 0,39 puntos por encima del propio sujeto, con d de Cohen 1,07 y 0,46; familia, amistad y trabajo alteran esas dos dimensiones.

La aportación valiosa no es demostrar una personalidad interna, sino mostrar que un perfil declarado puede no coincidir con la conducta generada. El caso cualitativo es ilustrativo: un sujeto instruido como muy desagradable se autocalifica en 1,7, pero produce un diálogo más bien cauteloso y cooperativo y los observadores le asignan alrededor de 3. El diseño amplía el self-report con evidencia conductual, prueba tres familias de modelos, cuatro variantes de prompt, orden inverso y batch, y añade una comparación humana con acuerdo entre evaluadores. Es una dirección metodológica prometedora para medir fidelidad conductual de personas inducidas.

No obstante, los «informantes» no son observadores externos independientes: en la condición principal son instancias del mismo GPT-4o que genera perfiles, relaciones, escenarios, ambos lados del diálogo y la puntuación IPIP. Observadores y humanos ven los diálogos, mientras el self-report recibe directamente los marcadores de personalidad; esa asimetría favorece estructuralmente la coincidencia observador–humano y la coincidencia self-report–prompt. El perfil inyectado no es ground truth psicológico, promediar agentes correlacionados no establece fiabilidad y la meseta de cinco a siete no se valida con intervalos, remuestreo ni conjunto externo. La evidencia humana comprende 16 casos con dos evaluadores por caso; el checklist, sin embargo, habla de 12 participantes y confirma que no hubo aprobación o exención de un comité ético. El repositorio que la versión final declara disponible responde 404, de modo que prompts, outputs y análisis no son actualmente reproducibles. La conclusión defendible es que informes multiagente capturan mejor la impresión humana de estas conversaciones sintéticas bajo este protocolo, no que midan una personalidad verdadera o más objetiva del LLM.

Research question

Can an aggregated Big Five evaluation from multiple observer agents, after conversing with a subject agent in different relational contexts, better reflect its simulated behavior and human judgment than the direct self-report of the LLM itself?

Method

Multiagent simulation experiment. Each subject receives a name, age, gender, and a Big Five vector of five integer levels 1–6 converted into three adjectival markers per trait. For 100 subjects, 15 observers are assigned, five family members, five friends, and five work relationships, and GPT-4o generates five scenarios per pair. Subject and observer alternate dialogue until agreeing to end. The subject completes 50 IPIP items about itself; each observer completes the same items about the subject using the dialogues, and the collective report is the mean of N observers. Spearman correlations are calculated with the injected profile, self-report, and human ratings; deviations, paired t tests, and Cohen's d; curves with 1–15 observers; relationship effects; and sensitivity to model, prompt style, scale order, and batch. Simulation at temperature 1, questionnaires at 0. The editorial review read the 16 published pages, tables, figures, prompts, and checklist, and verified the DOI and linked artifact.

Sample: The main experiment uses 100 subject instances of the same GPT-4o, not 100 independent models or persons. Each is combined with 15 observer instances and five scenarios per pair, that is, 75 potential synthetic conversations per subject before aggregating ratings. Sensitivity analyses include Qwen2.5-72B-Instruct and Llama-3-70B-Instruct. The human evaluation publishes 16 subject cases with two raters per case; participants are native English speakers residing in the United Kingdom, the United States, New Zealand, Canada, or Australia, take approximately 15 minutes, and receive 2.25 GBP. The manuscript does not clarify in a reconcilable manner how many unique persons participated: the checklist declares 12 participants, while the appendix describes 16 cases, two annotators per case, and five dialogues per participant.

Findings

  • The self-report almost exactly follows the injected profile: Spearman correlations of 0.97 in openness, 0.95 in conscientiousness and extraversion, 0.94 in agreeableness, and 0.93 in neuroticism.
  • The aggregated observer report correlates less with the injected profile: 0.55, 0.85, 0.84, 0.84, and 0.86 respectively, which evidences a gap between direct instruction and dialogued behavior.
  • Compared with humans, the observer improves openness from −0.25 to 0.48, agreeableness from 0.63 to 0.85, and neuroticism from 0.22 to 0.42; the self-report remains slightly above in conscientiousness and extraversion in the main table.
  • In the appendix, the absolute distance to human ratings is smaller for observers than for self-report across the five dimensions and in GPT-4o, Qwen2.5, and Llama-3.
  • Agreement between the two human raters is heterogeneous: Pearson 0.31 in openness, 0.69 in conscientiousness, 0.73 in extraversion, 0.59 in agreeableness, and 0.45 in neuroticism.
  • The aggregated correlation usually grows and stabilizes approximately upon incorporating five to seven observers; openness remains lower, around 0.60–0.65.
  • Observers score agreeableness 0.91 points above the self-report, with d = 1.07, and conscientiousness 0.39 points above, with d = 0.46; both differences are declared significant at p < 0.05.
  • No significant deviations of self-report versus observer are identified in openness, extraversion, or neuroticism for GPT-4o in the main condition.
  • Work versus family/friendship changes conscientiousness, and family versus work changes agreeableness; the other three dimensions show no significant differences between contexts.
  • Qwen2.5 and Llama-3 reproduce the positive deviation of agreeableness and, to a lesser extent, conscientiousness; Qwen2.5 adds a significant deviation in openness.
  • The neutral, inverted-scale, and batch variants maintain the general deviation pattern, although magnitudes change and some significances are not preserved.
  • The qualitative case shows that a minimum-agreeableness prompt produces a self-report of 1.7, but the dialogue preserves cooperation and the aggregated observers assign 3.0.

Limitations

  • The latent profile is a written instruction, not an observed psychological variable or independent ground truth; it measures the personality that was asked to be simulated.
  • The self-report explicitly receives the markers that it later scores, so its almost perfect correlation with the injected profile is partly a mechanical consequence of the design.
  • Observers and humans evaluate the same behavioral material, while the self-report evaluates an internal instruction without those dialogues; the comparison confronts different information sources.
  • GPT-4o generates relationships and scenarios and acts as subject, interlocutor, and judge in the main experiment, creating circularity and shared biases.
  • The observers are not independent models, trainings, or persons: they are correlated instances of the same system. The hypothesis that random errors cancel out is not proven.
  • The observer participates in the dialogue that it later judges; its own responses can shape the subject's behavior and its subsequent evaluation.
  • The use of IPIP by another LLM does not eliminate possible contamination from knowledge of the test; it shifts the questionnaire from the subject to a judge-model that may also recognize it.
  • The relationships and scenarios are not assigned as equivalent stimuli across conditions. Differences between family, friendship, and work may arise from the generated content, not from the relational context alone.
  • The plateau of five to seven observers is identified visually on the same curves used to evaluate it, without a confidence interval, bootstrap, validation set, or formal comparison of cost and precision.
  • The analogy with Dunbar's number is post hoc and does not constitute evidence that synthetic agents share human social structure.
  • The human sample is small: 16 cases with two raters. Per-trait correlations with that n are unstable and are not accompanied by confidence intervals.
  • Human agreement is low for openness and moderate for neuroticism, limiting the strength of the criterion used to declare validity.
  • No direct comparison with trained human coders, behavior classifiers, or observers from another family is presented that would allow isolating the specific value of the multiagent framework.
  • The t tests treat conditioned instances of the same model as observations; pseudoreplication, dependence by subject/observer/scenario, and a hierarchical model are not discussed.
  • Five traits and multiple relationships, models, and prompts are tested without correction for multiple comparisons; asterisks are shown, but not exact p values or intervals.
  • Cohen's d is described with pooled standard deviation despite the paired design, without precisely specifying the variant of the estimator.
  • The claim that alignment training causes the agreeableness and conscientiousness biases is speculative: no training is manipulated and no pre/post alignment checkpoints are compared.
  • Generated names, ages, gender, relationships, and scenarios may introduce stereotypes, but demographic effects, content, or generator bias are not audited.
  • The limitations section acknowledges simplified scenarios and relationships and discrepancies between scales, but does not analyze risks of applying these ratings in mental health, education, or employment.
  • The checklist declares that there was no ethics-board approval or exemption determination; its figure of 12 participants does not match the appendix description of 16 cases and two annotators per case.
  • The published version links a GitHub repository as available code, but the URL returns 404 and no identifiable replacement repository was found; data, seeds, scripts, or exclusions cannot be audited.
  • The results are based on GPT-4o, Qwen2.5-72B-Instruct, and Llama-3-70B-Instruct with 2024–2025 configurations and do not automatically generalize to later versions.
  • The published cost, approximately 2.9 USD per GPT-4o subject, does not include a total accounting of tokens, exact service date/snapshot, or variance across runs.

What the study does not establish

  • It does not demonstrate that an LLM possesses an internal, stable personality equivalent to human personality.
  • It does not demonstrate that the observer report is objective; it shows greater coincidence with a small sample of human impressions on the same dialogues.
  • It does not demonstrate construct, factorial, criterion, or predictive validity of the IPIP applied to LLM agents.
  • It does not demonstrate that the injected profile is the agent's true personality or that deviating from it is necessarily an error.
  • It does not demonstrate that five to seven is a universal optimal number of observers; it is a descriptive plateau of this protocol.
  • It does not demonstrate a collective wisdom effect among independent sources, because the instances share model, prompts, and generator.
  • It does not causally establish that alignment training produces the observed deviations.
  • It does not prove that family, friendship, or work cause perception changes separate from the scenarios and dialogues generated for each condition.
  • It does not show that the scores predict satisfaction, safety, interaction quality, or human outcomes in real applications.
  • It does not validate its use for selection, diagnosis, education, mental health, relationships, or high-impact decisions.
  • It does not allow generalization to all LLMs, languages, cultures, architectures, or longitudinal interactions.
  • It does not currently offer an independent reproduction of the results because the linked public artifact is not available.

Traceability

Scope: Full text

Version: Findings of ACL: EMNLP 2025, pp. 21086–21101, DOI 10.18653/v1/2025.findings-emnlp.1150; 16-page published PDF and Responsible NLP checklist reviewed; referenced GitHub artifact returned HTTP 404 on 15 Jul 2026

Consulted source: https://aclanthology.org/2025.findings-emnlp.1150.pdf

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

  • OpenAI GPT-4o (main subject, observer, relationship and scenario generator)
  • Qwen/Qwen2.5-72B-Instruct (appendix sensitivity analysis)
  • meta-llama/Meta-Llama-3-70B-Instruct (appendix sensitivity analysis)

Instruments and metrics

  • 50-item International Personality Item Pool (IPIP) Big Five questionnaire
  • Five-point Likert self-report
  • Five-point Likert individual observer-report
  • Mean aggregated multi-observer report
  • Injected Big Five profile with three bipolar personality markers per dimension
  • Spearman rank correlation
  • Paired-samples t test
  • Cohen's d
  • Absolute difference from human ratings
  • Pearson inter-rater correlation for human ratings

Data used

  • Synthetic agent profiles: common US names, random age 15–80, gender and injected Big Five levels
  • 100 prompt-conditioned subject agents in the main experiment
  • 15 observer agents per subject: five family, five friend and five workplace relations
  • Five GPT-4o-generated scenarios per subject–observer pair
  • Self-report and observer-report IPIP ratings
  • Human judgment set: 16 subject cases, two annotators per case
  • Prompt-sensitivity conditions: default, neutral, reversed scale and 50-item batch

Evidence and location

  • Final publication, DOI, objective, and informant reports framework: Findings of ACL: EMNLP 2025, pp. 21086–21087, abstract and sections 1–2
  • Profiles, contexts, scenario generation, and three types of report: Published paper, sections 3.1–3.3, pp. 21088–21089
  • 100 subjects, 15 observers, five scenarios, IPIP-50, and models: Published paper, section 4, p. 21090; Appendix A.3–A.4, pp. 21098–21100
  • Correlations with latent profile and human ratings: Published paper, Table 1 and section 5.1, pp. 21090–21091; Appendix Tables 5–6, pp. 21100–21101
  • Descriptive effect of 5–7 observers and analogy with Dunbar: Published paper, section 5.2 and Figure 3, pp. 21091–21092
  • Agreeableness and conscientiousness deviations, t tests, and Cohen's d: Published paper, Table 2 and section 5.3, pp. 21092–21093
  • Relational context effect and speculation about alignment: Published paper, section 5.4 and Figure 4, p. 21093
  • Case of disagreement between prompt, self-report, and behavior: Published paper, section 5.5 and Tables 3–4, pp. 21093–21095
  • Model/prompt variations, temperature, infrastructure, and cost: Published paper, Appendix A.3–A.4 and Figure 5, pp. 21098–21100
  • 16 cases, two annotators, recruitment, compensation, and human agreement: Published paper, Appendix A.5, Tables 5–7, pp. 21100–21101
  • Stated limitations of scenarios, relationships, and scales: Published paper, Limitations, p. 21095
  • Absence of ethics review and sample discrepancy: Responsible NLP Checklist for 2025.findings-emnlp.1150, item D4 reports 12 participants and no ethics-board approval/exemption; Appendix A.5 reports 16 cases with two annotators each
  • Public artifact unavailable: Published paper footnote 1 links github.com/leslie071564/llm_personality_observer; GitHub API returned HTTP 404 and repository/code searches found no replacement on 15 Jul 2026