Evaluation Drift in LLM Personality Induction: Are We Moving the Goalpost?

Trait induction and control2026arXivApproved editorial review

Authors: Prateek Rajput, Yewei Song, Iyiola E. Olatunji, Jacques Klein, Tegawendé F. Bissyandé

Keywords: Personality, Persona conditioning, Psychometrics, Evaluation validity, Reproducibility, Data privacy

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

5
Authors
16
Findings
23
Limitations
5
Evidence

Editorial summary

English

This preprint, listed on an author's page as accepted and forthcoming at LREC 2026, raises a useful concern: evaluating personality induction one trait at a time can conceal failure on the complete Big Five vector. It trains models on human essays labeled with the 32 binary OCEAN combinations and compares SFT, DPO, and ORPO using IPIP-NEO. Its tables report very low exact match, at most 3 of 32 profiles. However, the implementation does not support the paper's two strongest explanations: it neither demonstrates that fine-tuning stabilizes responses under prompt rephrasing nor that essays lack sufficient personality cues.

The described corpus has 2,467 essays, 1.9 million words, and about 770 words per text. The repository releases 2,066 unique moderated essays split 1,652/207/207 for train/validation/test. SFT receives the OCEAN vector and learns to generate the corresponding human essay; DPO/ORPO contrast it with three essays whose complete vector differs. Another condition appends synthetic responses to about half of IPIP-NEO. Reported models are Gemma-2-2B, Llama-3.2-3B, Gemma-7B, Llama-3.1-8B, and GPT-3.5-turbo-0125, plus unidentified “uncensored” counterparts.

The central problem affects RQ1. The code does not calculate variation for the same profile across S1, S2, and S3. It calculates the standard deviation of scores across 32 different target profiles inside each file. This dispersion measures sensitivity to the requested profile, not to rephrasing. A collapsed model that always answers identically would obtain zero deviation, the apparent optimum. The reported 15%-33% reduction can therefore mean weaker personality differentiation rather than stronger stability; the released statistic does not measure the rephrasing-robustness claim.

RQ2 also contains target leakage. After generating an essay, every questionnaire item again says “Answer the question as if you are” followed by the full target OCEAN vector. The essay is added as context, but the explicit label remains. Exact match mixes compliance with that instruction, essay effects, and self-report formatting; it does not infer a vector from an unguided essay. A 1/32 reference is reasonable for the artificially balanced 32 vectors, but each configuration has one pass. For 3/32, a binomial reference with p=1/32 gives P(X≥3)≈0.077 and a Wilson 95% interval of about 3.24%-24.22%, overlapping 0.55%-15.74% for 1/32. Moreover, 3/32 is the maximum across 36 configurations. There are no repeats, seeds, intervals, tests, multiplicity correction, or power analysis.

The safety conclusion uses only counts from zero to three between censored and “uncensored” models, without exact identities, snapshots, architectural matching, or an equivalence test. “No significant gains” does not establish no effect or remove alignment as a confounder. The two qualitative cases are not causal explanations either; the first quoted essay appears verbatim in the SFT training split.

Data and ethics claims conflict with the artifact. Moving from 2,467 to 2,066 leaves 401 records, not “approximately 300,” without a manifest. The script filters hate, harassment, self-harm, sexual, and violence categories; it does not detect PII although the paper attributes filtering to personal data. Essays contain names, locations, relationships, and health narratives; a conservative check finds 11 with “@,” nine mentioning a therapist, and at least 63 with selected health terms. No consent, lawful-basis, or re-identification audit supports the no-PII and GDPR claims. The paper assigns Apache-2.0 to the dataset without upstream evidence in the repository and promises MIT while shipping no LICENSE.

The questionnaire condition adds noise: it adjusts raw responses before reverse scoring. Of 10,330 randomized synthetic trait blocks, 210 fall in the wrong class under the study's threshold; 78 negative blocks land exactly at 3.0 and are classified positive. SFT train is also imbalanced, with 24-131 examples per vector.

The repository makes preprocessing, SFT/DPO/ORPO for four Hugging Face families, local inference, and processed data inspectable; all Python files compile syntactically. Missing components include questions.json, original/moderated CSVs, uncensored models, the GPT-3.5 pipeline, questionnaire DPO/ORPO data, checkpoints, raw responses and results, logs, seeds, tests, CI, lockfile, and container. The main text specifies DPO/ORPO rank 8, dropout 0.1, cosine scheduling, and 10% warmup; code and appendix specify rank 16, dropout 0.05, linear scheduling, and 500 steps. The faithful conclusion is limited: the work descriptively finds few complete vectors and correctly emphasizes joint exact match, but its stability metric is misaligned, evaluation leaks the target, and the artifact cannot reproduce the numbers. It does not establish latent personality, causal insufficiency of essays, safety equivalence, or legal and ethical compliance.

Español

Este preprint, que una página de autor presenta como aceptado y pendiente de aparición en LREC 2026, plantea una crítica pertinente: evaluar la inducción de personalidad rasgo por rasgo puede ocultar el fracaso del vector Big Five completo. Entrena modelos con ensayos humanos etiquetados en 32 combinaciones OCEAN binarias y compara SFT, DPO y ORPO mediante IPIP-NEO. Sus tablas muestran exact match muy bajo, como máximo 3 de 32 perfiles. Sin embargo, la implementación no sostiene las dos explicaciones más fuertes del paper: ni demuestra que el fine-tuning estabilice respuestas frente a reformulaciones ni que los ensayos carezcan de señales suficientes de personalidad.

El corpus descrito tiene 2.467 ensayos, 1,9 millones de palabras y unas 770 palabras por texto. El repositorio libera 2.066 ensayos únicos tras moderación: 1.652/207/207 para train/validación/test. SFT recibe el vector OCEAN y aprende a generar el ensayo humano correspondiente; DPO/ORPO lo contrastan con tres ensayos cuyo vector no coincide por completo. Otra condición añade respuestas sintéticas a aproximadamente la mitad de IPIP-NEO. Se reportan Gemma-2-2B, Llama-3.2-3B, Gemma-7B, Llama-3.1-8B y GPT-3.5-turbo-0125, más contrapartes «uncensored» no identificadas.

El problema central afecta a RQ1. El código no calcula variación del mismo perfil entre S1, S2 y S3: calcula la desviación estándar de las puntuaciones entre 32 perfiles objetivo diferentes dentro de cada archivo. Esa dispersión mide sensibilidad al perfil solicitado, no al rephrasing. Un modelo colapsado que conteste siempre igual obtendría desviación cero, el supuesto mejor resultado. La reducción del 15%-33% puede significar menor diferenciación entre personalidades y no mayor estabilidad; el claim de robustez a reformulaciones no está medido por la estadística liberada.

RQ2 contiene además fuga de target. Después de generar un ensayo, cada ítem vuelve a decir «Answer the question as if you are» seguido del vector OCEAN completo. El ensayo se añade como contexto, pero la etiqueta explícita permanece. El exact match mezcla obediencia a esa instrucción, efecto del ensayo y formato de autorreporte; no infiere el vector desde un ensayo unguided. El 1/32 es razonable para los 32 vectores artificialmente balanceados, pero solo hay una pasada por configuración. Para 3/32, una referencia binomial con p=1/32 da P(X≥3)≈0,077 y un intervalo Wilson 95% de aproximadamente 3,24%-24,22%, solapado con 0,55%-15,74% para 1/32. Además, 3/32 es el máximo entre 36 configuraciones. No hay réplicas, semillas, intervalos, tests, multiplicidad ni potencia.

La conclusión sobre seguridad se basa solo en conteos de 0-3 entre modelos censurados y «uncensored», sin identidades exactas, snapshots, emparejamiento arquitectónico o test de equivalencia. «No significant gains» no prueba ausencia de efecto ni permite eliminar el alineamiento como confusor. Los dos casos cualitativos tampoco son explicaciones causales; el primer ensayo citado aparece literalmente en el split de entrenamiento SFT.

Los datos y la ética presentan contradicciones. De 2.467 a 2.066 faltan 401 registros, no «aproximadamente 300», sin manifest explicativo. El script filtra odio, acoso, autolesión, sexualidad y violencia; no detecta PII aunque el texto atribuye el filtrado a personal data. Los ensayos contienen nombres, ubicaciones, relaciones y salud; una comprobación conservadora encuentra 11 con «@», nueve que mencionan a un terapeuta y al menos 63 con términos de salud seleccionados. No hay auditoría de consentimiento, base jurídica o reidentificación que sustente «sin PII» y GDPR. El paper atribuye Apache-2.0 al dataset sin prueba upstream en el repo y promete MIT sin incluir LICENSE.

La condición con cuestionario también introduce ruido: ajusta respuestas crudas antes del reverse scoring. De 10.330 bloques sintéticos aleatorios, 210 quedan en la clase equivocada bajo el umbral del estudio; 78 bloques negativos caen exactamente en 3,0 y se clasifican positivos. El train SFT está desequilibrado, con 24-131 ejemplos por vector.

El repositorio permite inspeccionar preprocessing, SFT/DPO/ORPO de cuatro familias Hugging Face, inferencia local y datos procesados; todo el Python compila sintácticamente. Pero faltan questions.json, CSV original/moderado, modelos uncensored, pipeline GPT-3.5, DPO/ORPO con cuestionarios, checkpoints, respuestas y resultados brutos, logs, seeds, tests, CI, lockfile y contenedor. El texto principal da para DPO/ORPO rango 8, dropout 0,1, cosine y 10% warmup; código y apéndice dan rango 16, dropout 0,05, linear y 500 pasos. La conclusión fiel es limitada: el trabajo muestra descriptivamente pocos vectores completos y recuerda que el exact match conjunto importa, pero su métrica de estabilidad está mal alineada, la evaluación filtra el target y el artefacto no reproduce las cifras. No establece personalidad latente, insuficiencia causal de los ensayos, equivalencia de seguridad ni cumplimiento legal o ético.

Research question

Does post-training reduce the sensitivity of IPIP-NEO responses to three prompt formats, can SFT/DPO/ORPO jointly induce the five OCEAN traits from human essays, and does that performance change between aligned models and "uncensored" counterparts?

Method

Human essays are filtered with five binary labels and models are trained to generate text conditioned on the OCEAN vector via SFT, DPO, or ORPO, with or without synthetic responses from one half of IPIP-NEO. In evaluation, the 32 vectors are enumerated, an essay is generated, and items from the other half are answered in S1/S2/S3. The audit reviewed 14 pages, TeX, tables, prompts, 2,066 base essays, 4,132 augmented records, 6,198 preference pairs, and code at commit e32a7ad8da2cf055fb4e4c52f39ac1bc451a6b5b.

Sample: The artifact releases 2,066 unique essays after filtering 401 with respect to the 2,467 declared. The SFT split is 1,652/207/207 and its 32 joint profiles are imbalanced: 24-131 examples per profile in train. Each configuration is evaluated once over 32 enumerated vectors; Table 5 contains 36 non-missing configurations. There are no replicates, training seeds, or human sample.

Findings

  • The maximum exact match reported is 3/32; the majority obtain 0-2.
  • The work is correct in requiring joint evaluation of five traits.
  • RQ1 uses dispersion across profiles, not variation of the same profile across prompts.
  • A collapsed model would have zero deviation.
  • Each question reveals the full target OCEAN vector.
  • The evaluation does not isolate the information contributed by the essay.
  • 3/32 is not conventionally distinguishable from p=1/32 without multiplicity.
  • No censored/uncensored equivalence is demonstrated.
  • The first qualitative case is in the training split.
  • The counts imply 401 filtered essays, not approximately 300.
  • Content moderation is not a PII filter.
  • 210 of 10,330 randomized blocks contradict their label after scoring.
  • The local open-source code is inspectable and compiles.
  • GPT-3.5, uncensored, questions.json, variants, and raw results are missing.
  • Main text and appendix/code give incompatible hyperparameters.
  • There is no MIT license or upstream evidence of Apache-2.0.

Limitations

  • LREC only verified as to appear on author page.
  • Central metric misaligned with RQ1.
  • Complete target leakage in evaluation.
  • One pass over 32 vectors per configuration.
  • No replicates, seeds, intervals, tests, or multiplicity.
  • NaN 6%-10% excluded from means.
  • Threshold 3.0 turns exact neutrality into positive.
  • Imbalanced training profiles.
  • Uncensored not identified or matched.
  • No equivalence or power test for RQ3.
  • Cases without causal attribution; one is in train.
  • 401 records not reconciled.
  • Moderation does not eliminate or audit PII.
  • No consent, legal basis, or reidentification.
  • Apache-2.0 and MIT licenses not substantiated.
  • 210 misaligned synthetic blocks.
  • questions.json and source CSV absent.
  • No DPO/ORPO with questionnaires.
  • No GPT-3.5 or uncensored pipeline.
  • No checkpoints, outputs, logs, or manifests.
  • Incompatible hyperparameters.
  • Seed does not control pandas.sample.
  • No tests, CI, lockfile, or container.

What the study does not establish

  • Stability to reformulations.
  • Latent or human personality in the models.
  • That the essays lack OCEAN signals.
  • Profile inference from the essay alone.
  • Significant improvement across methods.
  • Censored/uncensored equivalence.
  • Absence of safety alignment effect.
  • Token-level causes of errors.
  • Generalization to persons or texts.
  • Psychometric validity of LLM self-report.
  • Anonymity, consent, or GDPR.
  • Verifiable license of the corpus or repo.
  • Reproduction of figures with a clean clone.

Traceability

Scope: Full text

Version: arXiv:2605.16996v1, submitted 2026-05-16, 14 pages, complete TeX; author page lists LREC 2026 to appear; repository audited at e32a7ad8da2cf055fb4e4c52f39ac1bc451a6b5b

Consulted source: https://arxiv.org/abs/2605.16996v1

Review: Codex 14-page visual full-text, complete TeX, metric, target-leakage, statistics, missingness, data-quality, privacy, licensing, repository code and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Gemma-2-2B
  • Llama-3.2-3B
  • Gemma-7B
  • Llama-3.1-8B
  • GPT-3.5-turbo-0125
  • Unspecified uncensored Gemma-2-2B counterpart
  • Unspecified uncensored Llama-3.2-3B counterpart
  • Unspecified uncensored Llama-3.1-8B counterpart

Instruments and metrics

  • IPIP-NEO with unreleased train/test item split
  • S1 numeric, S2 verbal and S3 alphabetical prompts
  • Reverse-scored trait mean with threshold 3.0
  • Five-dimensional exact match over 32 binary vectors
  • Per-trait binary accuracy
  • Standard deviation across target profiles, presented as prompt variance
  • NaN exclusion for unparsable responses
  • Independent metric, leakage, statistics, data, privacy, licensing, code and reproducibility audit

Data used

  • Essays Dataset reported as 2,467 essays and 1.9 million words
  • Released SFT essays: 1,652 train, 207 validation, 207 test
  • Released questionnaire SFT records: 3,304 train, 414 validation, 414 test
  • Released preference pairs: 5,577 train, 621 test
  • Claimed but unreleased questionnaire preference pairs: 11,154 train, 1,242 test
  • Unreleased questions.json and exact IPIP-NEO item split
  • Unreleased raw model responses and evaluation results

Evidence and location

  • Text, tables, prompts, ethics, and appendix: arXiv:2605.16996v1; PDF sha256 326793423925479571bb91a97751ba78af43ec2b0cde54cae13cafd9b3761e8c; TeX sha256 450a4927d1acabc4ae7bcf96d901de256f4fba044d1ca1f1858cb33c6381a823
  • Editorial status: jacquesklein2302.github.io/publications lists LREC 2026 to appear; DBLP lists CoRR abs/2605.16996 at audit time
  • Metric, prompts, scoring, and artifact: github.com/pkrajput/personality_induction at e32a7ad8da2cf055fb4e4c52f39ac1bc451a6b5b
  • Counts, privacy, and noise: Released data audit: 2,066 base essays, 10,330 randomized trait blocks, 210 threshold contradictions
  • Independent audit: reports/verification/article-329-evaluation-drift-variance-metric-target-leakage-statistics-data-privacy-code-and-reproducibility-audit.json