Maintaining Stable Personas? Examining Temporal Stability in LLM-Based Human Simulation

Personas, identity, and agents2026arXivApproved editorial review

Original title: Stable Personas: Dual-Assessment of Temporal Stability in LLM-Based Human Simulation

Authors: Jana Gonnermann-Müller, Jennifer Haase, Nicolas Leins, Thomas Kosch, Sebastian Pokutta

Keywords: Large Language Models, Personality, Persona, Personality Control, AI Safety

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

5
Authors
20
Findings
89
Limitations
14
Evidence

Editorial summary

English

The definitive publication is titled “Maintaining Stable Personas? Examining Temporal Stability in LLM-Based Human Simulation,” appeared in CHI EA 2026, and has DOI 10.1145/3772363.3799334. This review audits all six pages and the expanded 19-page methodological manuscript arXiv:2601.22812v2, revised 20 May 2026. The old title “Stable Personas: Dual-Assessment…” is retained only as the original title. No official public repository, dataset, or code was found; the arXiv source contains LaTeX and figures but not outputs, scripts, or an executable configuration.

The study does not test a complete “personality.” It constructs four ADHD symptom-intensity conditions: high, moderate, low, and default without a persona. Text prompts turn DSM/ICD criteria into descriptions using “frequently,” “sometimes,” or “rarely”; a scale prompt fixes inattention, hyperactivity, and impulsivity at 6/7, 3/7, or 1/7; a third variant paraphrases the text. These are explicit instructions to portray symptom frequency, not people with histories, ages, contexts, strengths, impairment, subtypes, comorbidities, or real diversity. Semantic equivalence among the three prompts is not validated, and their absolute results differ substantially.

Seven models are tested, Claude Sonnet 4.5, DeepSeek V3.2, GPT-5.1, GPT OSS 120B, Gemini 3 Pro, Grok 4.1, and Llama 3.3 70B, with provider defaults. Commercial models use official APIs and open-weight models use Ollama. Immutable IDs, per-call dates, temperature/top-p/top-k, quantization, Ollama version, seeds, retries, and code are not released. Defaults are intended to represent typical usage, but they conflate architecture, alignment, provider, and decoding, and make replication impossible when aliases or defaults change.

Experiment I generates a first-person workday narrative and a 12-item CAARS ADHD Index self-report. It targets 3,500 runs and retains 3,473 after .77% attrition. Three LLMs, Claude, GPT-5.1, and Gemini, rate each narrative using the observer form. The definitive paper reports ordered separation: low Self M=1.22, Obs M=.56; moderate 18.1/15.5; high 29.1/20.3. The default condition is revealing: Self M=14.7 while Observer M=2.16. The same unprompted model describes itself as moderately symptomatic while observers see little symptom expression. This is not evidence of a stable internal representation; it is a dissociation between questionnaire compliance and generated behavior.

Apparent high/low stability is aided by floor and ceiling effects. Low is near zero and high near the maximum 36, leaving less room to vary; moderate and default occupy the center and have larger SDs. CHI partly acknowledges this. Persona's 92.3% self-report and 89.5% observer variance shares are not pure stability metrics either: the denominator includes deliberately huge low-versus-high prompt differences, making model, prompt, and residual shares look small. Yet the supplement shows large practical differences: moderate self-report ranges from 15.1 with the scale prompt to 21.1 with the text prompt; Claude's moderate observer mean is 10.0 versus 18.6 for GPT-5.1. A global share below 1% does not make prompts or models interchangeable.

Experiment II targets 1,400 conversations and retains 1,370; each lasts 18 turns and is assessed at turns 6, 12, and 18. Self-report barely changes: high +.2, moderate 0, low +.18, and default +.4 on a 36-point scale. Observer ratings fall from 17.5 to 14.0 for high and 13.2 to 10.8 for moderate; low and default rise slightly. The decline is descriptive: no test or interval is reported for a persona-by-turn interaction. The mixed model treats turn as a random factor with only three levels and reports 1.32% global variance; this average can conceal opposite directions across intensities and does not estimate a temporal slope.

Observer measurement also compares non-equivalent windows. At each checkpoint judges rate “the accumulated turns”: turn 12 contains 1–12 and turn 18 contains 1–18. Measurements are nested and share all earlier text. Adding neutral conversation can dilute symptom-marker density or alter judge heuristics on longer text even if later behavior has not drifted. Inter-rater reliability falls from ICC .83 at turn 6 to .75 and .69, reinforcing that the rating task changes with transcript length. Demonstrating drift would require non-overlapping windows, turn-level coding, and a slope/interaction model with uncertainty.

Human validation appears in the arXiv supplement, not the definitive extended abstract. Five master's-level psychologists rate only twenty narratives: six high, six moderate, six low, and two default, from Claude and Gemini under two prompt formats. Aggregated human–LLM ICC is .95, but its CI [.57, .99] is very wide and extreme group separation can inflate agreement. It does not validate the other five models, the paraphrased prompt, 18-turn conversations, drift sensitivity, or similarity to adults with ADHD. Humans and LLMs may agree on explicit stereotypes without the texts reproducing human distributions, heterogeneity, or behavior.

The “multi-informant” analogy is limited. In clinical psychology, self and informant provide partly independent observations about a real person. Here self-report is another output from a model that knows the prompt, while observer report is an LLM inference from synthetic text. Claude, GPT, and Gemini also act as both personas and judges in the study, allowing same-family and shared-bias effects. CAARS is validated for human adults, not as a generator metric. There is no human ground truth, patient dataset, diagnosis, or calibration of what score should correspond to low/moderate/high.

The statistical specification is insufficient for several claims. Model has seven levels, prompt three, and turn three, yet all are random effects; the sampled populations are not justified and component intervals are absent. The index is bounded and discrete without residual diagnostics. There is no preregistration, quantitative hypothesis, power analysis, correction for model/prompt comparisons, or missingness analysis. Overall attrition looks small but reaches 30% for Grok-default in Experiment II and 10% for GPT OSS-default in I. The supplement's N formulas are wrong: its Experiment I expression yields 4,550 rather than 3,500, and its written Experiment II expression yields 224 rather than 1,400. Cell tables do sum to the correct targets.

CHI and arXiv v2 also disagree on statistics for the same Experiment I and N=3,473. CHI reports low observer M=.56, SD=2.37 and moderate self M=18.1, SD=4.14; arXiv v2 reports .35/1.88 and 18.50/3.99. Default observer is 2.16/1.76 versus 2.19/1.75. No dataset or analysis version explains the change. This review prioritizes the definitive CHI figures and records the supplement values as a discrepancy rather than treating both as simultaneously exact.

The defensible conclusion is narrow: extreme frequency prompts make seven LLMs produce ordered and relatively repeatable CAARS scores, especially when directly asked about the symptoms they were instructed to portray. During neutral conversation, LLM judges perceive fewer high/moderate markers in longer cumulative prefixes. This usefully warns that self-report and observable expression are not interchangeable and that variability should be measured at multiple points. It does not demonstrate valid human simulation, realistic ADHD, an internal representation, personality stability, multi-agent behavior, clinical safety, or that reprompting fixes the issue.

Español

La publicación definitiva de este trabajo se titula “Maintaining Stable Personas? Examining Temporal Stability in LLM-Based Human Simulation”, apareció en CHI EA 2026 y tiene DOI 10.1145/3772363.3799334. Esta revisión audita sus seis páginas completas y el manuscrito metodológico ampliado arXiv:2601.22812v2, de 19 páginas, revisado el 20 de mayo de 2026. El título antiguo “Stable Personas: Dual-Assessment…” se conserva solo como título original. No se localizó repositorio, dataset ni código oficial público; el source de arXiv contiene LaTeX y figuras, pero no los outputs, scripts o configuración ejecutable.

El estudio no prueba una “personalidad” completa. Construye cuatro condiciones de intensidad de síntomas ADHD: high, moderate, low y default sin persona. Los prompts textuales convierten criterios DSM/ICD en descripciones con “frequently”, “sometimes” o “rarely”; el prompt de escala fija inattention, hyperactivity e impulsivity en 6/7, 3/7 o 1/7; una tercera variante parafrasea el texto. Son instrucciones explícitas para representar frecuencia de síntomas, no perfiles de personas con historia, edad, contexto, fortalezas, impairment, subtipo, comorbilidad o diversidad real. La afirmación de que las tres variantes son semánticamente equivalentes no se valida y sus resultados absolutos difieren sustancialmente.

Se prueban siete modelos, Claude Sonnet 4.5, DeepSeek V3.2, GPT-5.1, GPT OSS 120B, Gemini 3 Pro, Grok 4.1 y Llama 3.3 70B, con defaults de cada proveedor. Los modelos comerciales usan APIs oficiales y los open-weight, Ollama. No se publican IDs inmutables, fecha exacta por llamada, temperatura/top-p/top-k, cuantización, versión de Ollama, seeds, retries ni código. Usar defaults busca representar una experiencia típica, pero confunde arquitectura, alineamiento, proveedor y decoding; además impide repetir el experimento cuando cambien aliases y valores por defecto.

El Experimento I genera una narración en primera persona de un día laboral y un self-report de 12 ítems del CAARS ADHD Index. Apunta a 3.500 ejecuciones y conserva 3.473 después de 0,77% de attrition. Tres LLM, Claude, GPT-5.1 y Gemini, califican cada narración con la forma observer. La publicación informa una separación ordenada: low Self M=1,22, Obs M=0,56; moderate 18,1/15,5; high 29,1/20,3. El default, sin prompt, es revelador: Self M=14,7 pero Observer M=2,16. Es decir, el mismo modelo se autodescribe con síntomas moderados aunque su texto apenas los muestra a observadores. Esto no es evidencia de una representación interna estable; es una disociación entre obediencia al cuestionario y expresión textual.

La aparente estabilidad de high y low está favorecida por suelo y techo. Low queda casi en cero y high cerca del máximo 36, por lo que ambos tienen menos espacio para variar; moderate y default ocupan el centro y presentan SD mayores. Los autores lo reconocen parcialmente en CHI. La varianza atribuida a persona, 92,3% del self-report y 89,5% del observer, tampoco es una medida pura de estabilidad: el denominador incluye diferencias deliberadamente enormes entre prompts low y high. Ese contraste hace que modelo, prompt y residual parezcan pequeños. Aun así, el suplemento muestra diferencias prácticas grandes: en moderate, el self-report medio cambia de 15,1 con scale prompt a 21,1 con text prompt; el observer de Claude es 10,0 frente a 18,6 en GPT-5.1. “Menos del 1%” global no significa que prompt o modelo sean intercambiables.

El Experimento II apunta a 1.400 conversaciones y conserva 1.370; cada una dura 18 turnos y se evalúa en 6, 12 y 18. El self-report permanece casi constante: high +0,2, moderate 0, low +0,18 y default +0,4 sobre 36. En cambio, el observer baja high de 17,5 a 14,0 y moderate de 13,2 a 10,8; low y default suben ligeramente. Ese descenso es descriptivo: no se informa un test o intervalo para una interacción persona×turn. El mixed model trata turn como factor aleatorio con solo tres niveles y reporta 1,32% de varianza global; ese promedio puede ocultar direcciones opuestas por intensidad y no estima una pendiente temporal.

La medición observer tampoco compara ventanas equivalentes. En cada checkpoint, los jueces califican “the accumulated turns”: turn 12 incluye 1–12 y turn 18 incluye 1–18. Las tres observaciones están anidadas y comparten todo el texto previo. Añadir conversación neutral puede diluir la densidad de marcadores de síntomas o cambiar la heurística del juez sobre un texto más largo aunque la conducta posterior no derive. La fiabilidad entre jueces cae de ICC .83 en turn 6 a .75 y .69, reforzando que la tarea de evaluación cambia con la longitud. Para demostrar drift harían falta ventanas no solapadas, coding por turno y un modelo de slope/interaction con incertidumbre.

La validación humana aparece en el suplemento arXiv, no en el extended abstract definitivo. Cinco psicólogos con máster califican solo veinte narraciones: seis high, seis moderate, seis low y dos default, procedentes únicamente de Claude y Gemini y de dos formatos. El ICC humano–LLM agregado es .95, pero su CI [.57, .99] es muy amplio y la separación extrema de grupos puede inflarlo. No valida los otros cinco modelos, el prompt parafraseado, las conversaciones de 18 turnos, la sensibilidad al drift o la similitud con adultos con ADHD. Humanos y LLM pueden coincidir al reconocer los mismos estereotipos explícitos sin que los textos reproduzcan distribución, heterogeneidad o conducta humana.

La analogía “multi-informant” debe leerse con cautela. En psicología clínica, self e informant aportan observaciones parcialmente independientes sobre una persona real. Aquí el self-report es otra salida del modelo que conoce el prompt y el observer es un LLM que infiere los síntomas desde texto sintético. Claude, GPT y Gemini actúan además como persona y como jueces dentro del mismo estudio, permitiendo auto-familia y sesgos compartidos. CAARS está validado para adultos humanos, no como métrica de un generador. No hay ground truth humano, diagnóstico, dataset de pacientes ni calibración de qué score debería corresponder a low/moderate/high.

La especificación estadística es insuficiente para varias conclusiones. Modelo tiene siete niveles, prompt tres y turn tres, pero se tratan como random effects; no se justifican esas poblaciones ni se dan intervalos de componentes. El índice es discreto y acotado, sin diagnóstico de residuos. No hay preregistro, hipótesis cuantitativas, power, corrección por comparaciones de modelos/prompts o análisis de missingness. La attrition global parece baja, pero llega al 30% en Grok-default del Experimento II y 10% en GPT OSS-default del I, por lo que no es uniforme. Las fórmulas de N del suplemento son incorrectas: la de Exp I imprime 7×4×3×50+7×50=4.550, no 3.500; la de Exp II, tal como está escrita, da 224, no 1.400. Las tablas de celdas sí suman los targets correctos.

La publicación CHI y arXiv v2 también discrepan en estadísticas del mismo Experimento I y N=3.473. CHI da low observer M=0,56, SD=2,37 y moderate self M=18,1, SD=4,14; arXiv v2 da 0,35/1,88 y 18,50/3,99. Default observer es 2,16/1,76 frente a 2,19/1,75. No se documenta una versión de dataset o análisis que explique el cambio. Esta revisión prioriza las cifras definitivas de CHI y conserva las del suplemento como discrepancia, no como si ambas fueran simultáneamente exactas.

La conclusión defendible es estrecha: prompts de frecuencia extrema hacen que siete LLM produzcan scores CAARS ordenados y relativamente repetibles, sobre todo cuando se les pregunta directamente por los mismos síntomas. En conversación neutral, jueces LLM perciben menos marcadores high/moderate en prefijos acumulativos más largos. Esto es útil como alerta de que self-report y expresión observable no son intercambiables y de que se debe medir variabilidad en varios puntos. No demuestra simulación humana válida, ADHD realista, una representación interna, estabilidad de personalidad, comportamiento multi-agente, seguridad clínica ni que reprompting corrija el problema.

Research question

Do seven LLMs consistently repeat an instructed ADHD symptom intensity across independent generations and during an 18-turn conversation, according to CAARS self-report and observer ratings performed mainly by other LLMs?

Method

Two factorial experiments with seven models, three prompt formats and high/moderate/low/default conditions. Exp I: 3,473 workday narratives and CAARS self-reports, with three LLM judges. Exp II: 1,370 neutral conversations of 18 turns, evaluated cumulatively at 6/12/18. Means/SD are summarized and variance is decomposed with mixed linear models. The supplement adds five psychologists rating 20 narratives.

Sample: There is no sample of real simulated persons or adults with ADHD. The units are LLM generations: 3,473 in Exp I and 1,370 conversations in Exp II. The only human participation described in arXiv v2 is five psychologists with a master's degree who rate 20 synthetic texts; recruitment, compensation, selection of narratives, or ethical review are not detailed.

Findings

  • The definitive CHI EA 2026 publication uses a different title and DOI 10.1145/3772363.3799334.
  • Exp I retains 3,473 of 3,500 runs; Exp II retains 1,370 of 1,400 conversations.
  • The definitive scores are ordered low < moderate < high in self-report and observer.
  • In CHI, low has Self 1.22 and Observer 0.56; moderate 18.1/15.5; high 29.1/20.3.
  • Default without persona produces Self 14.7 but Observer 2.16, a large internal/observable dissociation.
  • High and low show smaller SD, partly compatible with ceiling and floor.
  • Moderate self-report differs by six points between scale prompt (15.1) and text prompt (21.1) in the supplement.
  • Moderate observer differs by 8.6 points between Claude (10.0) and GPT-5.1 (18.6).
  • The decomposition attributes 92.3%/89.5% to persona in Exp I, dominated by the low-high contrast.
  • Self-report changes only +0.2 high, 0 moderate, +0.18 low and +0.4 default between turn 6 and 18.
  • Observer drops 3.5 points in high and 2.4 in moderate, and rises slightly in low/default.
  • The global variance of turn is estimated at 0% self-report and 1.32% observer.
  • The judges rate cumulative prefixes, not independent six-turn windows.
  • LLM observer reliability falls from ICC .83 to .75 and .69 as the transcript grows.
  • Five psychologists and LLM judges agree on 20 narratives with aggregate ICC .95, CI [.57,.99].
  • Human validation only covers two target models, two prompts and 20 texts.
  • Attrition per cell reaches 30% in Grok-default Exp II and 10% in GPT OSS-default Exp I.
  • The written sample size formulas in arXiv v2 do not produce the declared targets.
  • CHI and arXiv v2 publish different means/SD for the same N without explaining the analytical version.
  • The robust result is stability of instructed responses; validity as human simulation is not evaluated.

Limitations

  • The definitive article is a six-page extended abstract and omits details from the supplement.
  • The arXiv supplement is formatted as an ICML submission and is not identical to the CHI version.
  • There is no public code, data, outputs, ratings, or executable environment.
  • There is no official repository located by title, arXiv ID, or authors.
  • Commercial models and their historical aliases cannot be reproduced.
  • Provider defaults are not listed numerically.
  • Temperature, top-p, top-k, max tokens, or seeds are not reported per model.
  • Retries, caching, concurrency, or independence control between runs are not reported.
  • Ollama has no documented version, quantization, hardware, or flags.
  • Architecture, alignment, provider, and decoding are confounded.
  • The persons are symptom intensities, not complete persons.
  • The prompts reduce ADHD to explicit frequency descriptors.
  • They do not include onset, impairment, multiple contexts, subtypes, or differential diagnosis.
  • They do not include strengths, coping, treatment, comorbidity, or neurodiverse heterogeneity.
  • Low ADHD is equivalent to a stereotypical absence of symptoms and not to a human distribution.
  • Default is not an equivalent control because it lacks the length and content of the persona prompt.
  • The three formats have no semantic equivalence validation.
  • The scale prompt is structurally different and contains explicit numerical anchors.
  • Mean differences by prompt contradict a simple practical equivalence.
  • CAARS is validated for human adults, not for LLM outputs.
  • There is no calibration between high/moderate/low and expected CAARS scores.
  • There is no human sample with ADHD or reference distribution.
  • Realism, representativeness, or clinical sensitivity/specificity is not measured.
  • Self-report asks the model about the same symptoms that the prompt orders it to represent.
  • Stable compliance with the questionnaire does not imply internal representation.
  • The term internal persona representation anthropomorphizes a conditioned output.
  • Moderate default self-report and almost-null observer question the interpretation of self-report.
  • High and low are close to ceiling/floor, restricting SD.
  • Absolute SD does not separate stability from floor/ceiling or incorrect calibration.
  • The variance of persona is inflated by deliberately constructed extremes.
  • A small residual percentage depends on the total range between conditions.
  • Small percentages of model/prompt hide large absolute differences.
  • No confidence intervals are given for variance components.
  • It is not explained how variance is attributed to the fixed effect persona.
  • The model has only seven levels as a random effect.
  • Prompt has only three levels as a random effect.
  • Turn has only three levels as a random effect.
  • Turn should be modeled as a trend/contrast to test drift.
  • The persona x turn interaction is not estimated or tested.
  • Declines of -3.5 and -2.4 lack CI or inferential test.
  • The global average of turn may cancel high/moderate downward with low/default upward.
  • The LMM formula does not explicitly model temporal slope per conversation or model.
  • Cumulative checkpoints share text and are not independent observations.
  • Turn 12 contains turns 1-12 and turn 18 contains 1-18.
  • Symptom density may dilute as neutral conversation is added.
  • Judges may change heuristics with input length.
  • Judge reliability decreases at each checkpoint.
  • Windows 1-6, 7-12, and 13-18 are not analyzed separately.
  • There is no per-turn coding of observable behaviors.
  • It is not clarified which backend generates the neutral conversation partner.
  • It is not clarified whether checkpoint self-reports remain in context and re-anchor the persona.
  • The neutral partner offers few natural opportunities to express hyperactivity/impulsivity.
  • A workday narrative and neutral chat do not cover transituational behavior.
  • Conversations last only 18 turns.
  • There is no interaction between several persons or agents with social roles.
  • Negotiation, trust, group, or opinion dynamics are not tested.
  • Three LLM judges are not independent informants of training/alignment.
  • Claude, GPT, and Gemini are simultaneously targets and evaluators.
  • Prevention of auto-family evaluation is not reported.
  • ICC(2,1) is single-rater reliability although three judges are aggregated.
  • ICC does not establish construct validity or human realism.
  • The N=20 human validation is very small.
  • There are only two default texts in human validation.
  • Only Claude and Gemini appear as target models in validation.
  • Only text and scale are validated, not paraphrase.
  • No long conversation or temporal change is validated.
  • The human-LLM CI [.57,.99] is very wide.
  • Artificial separation of intensities may inflate convergence ICC.
  • The five psychologists have a master's degree, but ADHD experience is not reported.
  • Random selection of the 20 narratives is not reported.
  • Recruitment, compensation, conflicts, or ethics of human raters are not reported.
  • There is no preregistration or prior analytical protocol.
  • There is no power analysis or precision target.
  • Residuals of a discrete 0-36 outcome are not diagnosed.
  • There is no robust/ordinal/binomial alternative analysis.
  • Descriptive comparisons by model and prompt are not corrected.
  • Missingness is attributed to API/JSON but it is not tested that it is random.
  • Global attrition hides 30% in one cell and 10% in another.
  • Exp I formulates N as 7x4x3x50+7x50, which gives 4,550.
  • The correct Exp I formula is 7x(3x3x50+50)=3,500.
  • Exp II omits x20 in the main product and its expression gives 224.
  • CHI/arXiv statistics differ without a data or analysis changelog.
  • The CHI abstract calls N=4,054 assessments without mentioning 12,201 observer ratings.
  • There are no published licenses or governance for outputs and ratings.
  • Cultural, gender, age, or ADHD presentation biases are not examined.
  • The harm of stereotyping a clinical condition in agents is not evaluated.
  • The reprompting recommendation is not tested experimentally.
  • CAARS stability does not demonstrate validity for prototype testing.
  • Generalization to personality, education, therapy, and social simulation is not proven.

What the study does not establish

  • It does not establish that LLMs validly simulate human persons.
  • It does not demonstrate that the texts reproduce adults with ADHD.
  • It does not demonstrate stability of general personality.
  • It does not demonstrate an internal representation or self-knowledge of the model.
  • It does not demonstrate that LLM self-report has the meaning of human self-report.
  • It does not demonstrate equivalence among the three prompts.
  • It does not demonstrate that model and prompt are irrelevant in practical magnitude.
  • It does not demonstrate behavioral drift separated from dilution and judge heuristics.
  • It does not demonstrate a significant or generalizable temporal slope.
  • It does not demonstrate stability beyond 18 turns.
  • It does not validate the seven models with humans.
  • It does not demonstrate stability in multi-agent social simulation.
  • It does not demonstrate clinical/educational safety or utility.
  • It does not test that periodic reprompting prevents observer decline.
  • It does not allow reproducing results from public artifacts.

Traceability

Scope: Full text

Version: Definitive CHI EA 2026 extended abstract, DOI 10.1145/3772363.3799334, 6 pages; audited with arXiv:2601.22812v2 methodological supplement, 19 pages

Consulted source: https://hcistudio.org/publication/gonnermann2026maintaining/gonnermann2026maintaining.pdf

Review: Codex definitive-CHI, full-text bilingual-fidelity, 6-page publisher visual, 19-page arXiv-v2 visual, DOI/title reconciliation, source-package, prompt-construct, clinical-measure, human-validation, LLM-judge, mixed-model, variance-denominator, floor-ceiling, longitudinal-window, attrition, arithmetic, cross-version consistency, reproducibility, ethics and multi-agent-claim audit; summaries written from the complete sources rather than abstract keyword extraction, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Anthropic Claude Sonnet 4.5, exact API snapshot not reported, persona and observer judge
  • DeepSeek V3.2, reported as 685B, exact API snapshot not reported, persona
  • OpenAI GPT-5.1, exact API snapshot not reported, persona and observer judge
  • OpenAI GPT OSS 120B via local Ollama, quantization and runtime version not reported, persona
  • Google Gemini 3 Pro, exact API snapshot not reported, persona and observer judge
  • xAI Grok 4.1, exact API snapshot not reported, persona
  • Meta Llama 3.3 70B via local Ollama, quantization and runtime version not reported, persona
  • A neutral conversation-partner LLM/prompt whose backend assignment is not clearly identified

Instruments and metrics

  • Three ADHD symptom-intensity prompt formats: text, 1–7 scale, and paraphrase
  • Default condition with no persona instruction
  • First-person typical-workday narrative task
  • Neutral open-question conversation-partner prompt
  • Conners Adult ADHD Rating Scales 12-item ADHD Index self-report, range 0–36
  • Parallel CAARS observer-report ratings from Claude, GPT-5.1, and Gemini
  • Standard deviation within experimental cells as the primary stability descriptor
  • Linear mixed-effects variance decomposition with persona fixed and model/prompt random
  • Conversation and turn random effects in Experiment II
  • Supplementary human annotation by five M.Sc.-level psychologists on 20 narratives

Data used

  • Experiment I target 3,500 runs; 3,473 retained after 27 failures
  • Experiment I: 3,473 self-reports and 10,419 LLM observer assessments
  • Experiment II target 1,400 conversations; 1,370 retained after 30 failures
  • Experiment II: 4,054 self-reports and 12,201 LLM observer assessments at three checkpoints
  • Twenty-narrative human-validation subset in arXiv v2: 6 high, 6 moderate, 6 low, 2 default, from Claude/Gemini and two prompts
  • Definitive six-page CHI EA 2026 paper, DOI 10.1145/3772363.3799334
  • Nineteen-page arXiv:2601.22812v2 methodological supplement and source package
  • No public raw conversations, questionnaire outputs, ratings, analysis code, model configs, or environment found as of 15 July 2026

Evidence and location

  • Definitive publication, title, DOI, and pages: CHI EA 2026 official author-hosted PDF, DOI 10.1145/3772363.3799334, pp. 1-6
  • Design, models, and prompts: Definitive CHI paper Sections 2-2.2 and Appendix Table 3; arXiv:2601.22812v2 Sections 2.1-2.3 and Appendix A.1-A.3
  • Definitive results Exp I: Definitive CHI paper Section 3 and Figure 1
  • Definitive results Exp II: Definitive CHI paper Section 3 and Figure 2
  • Variance and LMM specification: Definitive CHI paper Section 2.2 and Appendix Tables 1-2; arXiv v2 Equations 1-2
  • Counts and attrition per cell: arXiv:2601.22812v2 Appendix Tables 7-8, pp. 13-14
  • Incorrect N formulas: arXiv v2 Section 2.4; independent arithmetic versus Appendix Table 7 targets
  • Model/prompt absolute differences: arXiv v2 Section 3.1 and Appendix Tables 9-10
  • Cumulative checkpoints and declining ICC: arXiv v2 Sections 2.1 and 2.3; observer ICC .83/.75/.69
  • Supplementary human validation: arXiv v2 Section 2.3: five M.Sc. psychologists, 20 narratives, ICC and CIs
  • CHI versus arXiv v2 discrepancies: Definitive CHI Section 3 versus arXiv v2 Table 1 and Section 3.1 for the same N=3,473
  • Recognized limitations: Definitive CHI Section 4.1; arXiv v2 Section 4.1
  • Absence of repository/data/code: Official PDF, arXiv v2 source package, exact-title/arXiv-ID GitHub searches on 15 July 2026
  • Complete visual inspection: All 6 definitive CHI pages and all 19 arXiv v2 pages rendered and visually inspected on 15 July 2026