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.