ChildEval: When large language models meet children's personalities

Applications, bias, and safety2026ACL AnthologyApproved editorial review

Authors: Yanyan Luo, Xue Han, Chunxu Zhao, Ruiqiao Bai, Yaxing Zhang, Qian Hu, Lijun Mei, Junlan Feng

Keywords: Child preference following, Synthetic benchmark, Long-context personalization, Child safety

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

8
Authors
10
Findings
16
Limitations
5
Evidence

Editorial summary

English

ChildEval, published in Findings of ACL 2026, is a benchmark of synthetic child-preference following, not a psychological test of children's personalities despite its title. Its useful question is whether an LLM can remember and infer a preference for a synthetic 3-to-6-year-old profile when the preference is stated explicitly or expressed implicitly in a dialogue and then separated from the query by irrelevant sessions. The pipeline claims 29,000 Chinese personas generated with Qwen2.5-72B. Two persona-conditioned preferences per persona yield 58,000 candidates, of which FAISS semantic filtering retains about 46,000. The same model generates a query and a 6-to-10-exchange child-assistant dialogue for each preference; English is machine-translated from Chinese. Five areas and fourteen subtopics cover art, cognition, nutrition and activity, language, and social-emotional development. Real WildChat-1M user-LLM logs are inserted as distractors to create contexts up to 50 rounds or about 21,000 tokens. Qwen2.5-3B, Qwen3-4B, LLaMA3.1-8B, Mistral-7B, and DeepSeek-R1 are evaluated through prompting with and without persona, LoRA fine-tuning, and a Persona Steer Model. Preference Consistency derives from binary acknowledgement, violation, hallucination, and helpfulness judgments. Child-Oriented Evaluation adds Emotional Adaptation, Interaction Scaffolding, Developmental Appropriateness, and Engagement, also judged by an LLM. Reported results suggest that implicit preferences are usually harder than explicit statements and that irrelevant context often reduces consistency, although several curves are non-monotonic. Persona injection improves every selected five-distractor bar: Qwen3-4B rises from 78.7% to 89.1% on explicit data, while Qwen2.5-3B changes only from 75.7% to 75.8% on implicit data. Developmental Appropriateness is near ceiling for many models, while Interaction Scaffolding is much weaker, including 35.8% for Qwen2.5-3B without persona on explicit data. LoRA and PSM improve several metrics, especially scaffolding, but error analysis shows a trade-off: preference-unaware failures decrease while inconsistent or unhelpful responses increase. Human validation covers only 600 examples and raw agreement percentages. Qwen2.5-72B versus human reaches 88.83% overall PC and DeepSeek-R1 87.67%; COE dimensions reach 96.5-99.67%. This is not psychometric reliability. The paper does not report annotator count, recruitment, expertise, compensation, human instructions, label prevalence, human-human agreement, adjudication, kappa, or uncertainty. Near-99% agreement on binary criteria may be driven by class imbalance. Qwen2.5-72B also generates personas, preferences, dialogues, translations, and judge labels, creating stylistic dependence and circularity; DeepSeek is a second judge, not an independent human or behavioral criterion. The repository contains a substantial corpus, but the full audit changes the interpretation. Six files contain 45,757 rows and only 24,331 distinct Chinese personas. There are 45,755 English persona strings because repeated Chinese profiles are independently translated with variations. Although n-round always declares 6-10, 828 records have a different actual number of exchanges: 24 have one pair, one has two, and two have eleven. Index 27872 reverses roles in part of the English translation. Index 36947 declares ten science/technology rounds but contains one technology question and an unrelated answer about bear-shaped candy. The universal 6-10-turn claim is false for the released artifact, and prompt self-verification did not reliably enforce the schema. The stated 80/20 split is also non-reproducible: the script shuffles rows without a seed or persona grouping. Because 14,028 Chinese persona strings occur more than once, the expected split shares about 5,504 identities across train and test, roughly 22.6% of released personas, potentially inflating fine-tuning results. Safety requires particular caution. The paper correctly admits it does not model child-specific risk, developmental harm, or age-dependent safety. The README nevertheless says all data were manually checked to exclude harmful content. The corpus contains 5,770 allergy rows and hundreds involving asthma, ADHD, or other sensitive traits; 28 milk-allergy cases treat yogurt as safe or recommend it. A safety prompt exists but the evaluation code never calls it, and no manual-review protocol or labels are released. WildChat filtering, identifiers, possible minors, and personal-data handling are undocumented. Ethical disclosure is internally inconsistent: the checklist says there were no annotators or human subjects, while the paper reports 600 manually annotated examples, manual corpus review, and consented AI-assisted interaction summaries from volunteers about their children used to inform persona design. It also says no computational experiments were run and then claims experimental hyperparameters and descriptive statistics were reported. In fact, learning rate, batch size, epochs, optimizer, LoRA rank/alpha, PSM dimensions, seeds, decoding settings, repeated runs, intervals, and error bars are absent. The code release is partial: LoRA/PSM training, WildChat assembly, outputs, labels, and plotting are missing. The metric extractor raises NameError because four COE variables are undefined; the preference generator accidentally disables few-shot examples and can store a different topic from the one used in its prompt. No requirements, locked environment, tests, CI, release, or repository license are provided. The defensible contribution is a large Chinese-first synthetic diagnostic for explicit and implicit preference following under distractors, with enough released data for follow-up audits and informative method trade-offs. It does not demonstrate child personality, natural child language, developmental validity, learning, well-being, safety, caregiver benefit, performance with real children, or independent reproduction of training results and figures.

Español

ChildEval, publicado en Findings of ACL 2026, es un benchmark de seguimiento de preferencias infantiles sintéticas, no una prueba de personalidad psicológica de niños pese al título. Su pregunta útil es si un LLM puede recordar e inferir una preferencia de un perfil de 3 a 6 años cuando esa preferencia aparece de forma explícita en una frase o implícita en un diálogo y después queda separada de la consulta por sesiones irrelevantes. El pipeline declara 29.000 personas chinas generadas con Qwen2.5-72B. Se crean dos preferencias condicionadas por persona, 58.000 candidatas, y un filtrado semántico con FAISS retiene unas 46.000. Cada preferencia se asocia con una consulta y con un diálogo niño-asistente de 6 a 10 intercambios generado por el mismo modelo; el corpus inglés es una traducción automática del chino. Los temas cubren cinco áreas y catorce subtemas de arte, cognición, nutrición y actividad, lenguaje y desarrollo socioemocional. Para simular memoria larga, los autores intercalan como distractores conversaciones reales de WildChat-1M y prueban contextos de hasta 50 rondas o unos 21.000 tokens. Evalúan Qwen2.5-3B, Qwen3-4B, LLaMA3.1-8B, Mistral-7B y DeepSeek-R1 mediante prompting con y sin persona, ajuste LoRA y un Persona Steer Model que codifica el perfil e inyecta una representación mediante un adaptador y una compuerta. Preference Consistency combina juicios binarios de reconocimiento, violación, alucinación y utilidad. Child-Oriented Evaluation añade Emotional Adaptation, Interaction Scaffolding, Developmental Appropriateness y Engagement, todas puntuadas por otro LLM. Los resultados publicados indican que inferir preferencias implícitas suele ser más difícil que leerlas explícitamente y que añadir contexto irrelevante tiende a degradar el seguimiento, aunque varias curvas son no monótonas. Añadir la persona mejora todas las barras seleccionadas con cinco turnos distractores: el ejemplo mayor destacado es Qwen3-4B en explícito, de 78,7% a 89,1%, mientras Qwen2.5-3B apenas cambia en implícito, de 75,7% a 75,8%. Developmental Appropriateness queda cerca del techo en muchos modelos, pero Interaction Scaffolding es mucho menor, por ejemplo 35,8% para Qwen2.5-3B sin persona en el conjunto explícito. LoRA y PSM elevan varias métricas y el mayor avance descrito aparece en scaffolding, pero el análisis de errores muestra un intercambio real: bajan fallos por ignorar la preferencia y aparecen más respuestas inconsistentes o no útiles. La validación humana se limita a 600 ejemplos y porcentajes de acuerdo: Qwen2.5-72B frente a humano alcanza 88,83% en PC y DeepSeek-R1 87,67%; las cuatro dimensiones COE llegan a 96,5-99,67%. Esto no basta para hablar de fiabilidad psicométrica. No se publican número de anotadores, selección, experiencia, compensación, instrucciones humanas, prevalencia de etiquetas, acuerdo entre humanos, adjudicación, kappa ni incertidumbre. Un porcentaje de acuerdo cercano al 99% en variables binarias puede estar dominado por el desequilibrio de clases. Además, Qwen2.5-72B genera personas, preferencias, diálogos, traducciones y etiquetas de evaluación, creando dependencia estilística y circularidad; DeepSeek aporta un segundo judge, no un criterio humano o conductual independiente. La auditoría completa del repositorio confirma que sí hay un corpus sustancial, pero también descubre límites que cambian la lectura. Los seis archivos contienen 45.757 filas, no una tabla exacta de 46.000, y solo 24.331 personas chinas distintas. Hay 45.755 cadenas de persona inglesas porque cada repetición china se traduce de nuevo con variaciones. El campo n-round siempre declara 6-10, pero 828 registros no tienen ese número real de intercambios: 24 contienen una sola pareja de mensajes, uno contiene dos y dos contienen once; el índice 27872 invierte los roles en parte de la traducción inglesa. El índice 36947 declara diez rondas sobre ciencia y tecnología, pero solo contiene una pregunta sobre tecnología y una respuesta inconexa sobre caramelos con forma de oso. Por tanto, la afirmación universal de diálogos de 6-10 turnos no es cierta para el artefacto publicado, y la auto-verificación del prompt no funcionó de forma fiable. La separación 8:2 publicada tampoco es reproducible: el script baraja filas sin seed y sin agrupar por persona. Como 14.028 personas chinas aparecen en varias filas, el split esperado compartiría unas 5.504 identidades entre train y test, alrededor del 22,6% de las personas liberadas, lo que puede inflar el fine-tuning. El control de seguridad merece especial cautela. El paper reconoce correctamente que no modela riesgos infantiles, daño evolutivo ni seguridad dependiente de edad. Sin embargo, el README afirma que todo fue revisado manualmente para excluir contenido dañino. En el corpus hay 5.770 filas con alergias y cientos con asma, ADHD u otros rasgos sensibles; 28 casos describen alergia a la leche pero tratan el yogur como seguro o lo recomiendan. El repositorio incluye un prompt de seguridad, pero el script de evaluación nunca lo llama y no se publica protocolo, etiqueta ni rastro de la revisión manual. Tampoco se documenta cómo se filtran los logs reales de WildChat, si pueden incluir menores o datos personales, ni qué IDs se usaron. La divulgación ética es internamente inconsistente: el checklist marca que no hubo anotadores ni sujetos humanos, aunque el paper declara 600 ejemplos anotados, revisión humana del corpus y resúmenes consentidos de interacciones asistidas por IA aportados por voluntarios sobre sus hijos para diseñar las personas. También marca que no hubo experimentos computacionales y, a continuación, afirma haber informado de hiperparámetros y estadísticos descriptivos. En realidad no aparecen learning rate, batch, epochs, optimizer, rango/alpha de LoRA, dimensiones PSM, seeds, decoding, repeticiones, intervalos o barras de error. El código publicado es parcial: falta todo el entrenamiento LoRA/PSM, la construcción WildChat, outputs, labels y plotting. El extractor de métricas falla con NameError porque cuatro variables COE no están definidas; el generador de preferencias desactiva por error sus ejemplos few-shot y puede guardar un topic distinto del usado en el prompt. No hay requirements, entorno bloqueado, tests, CI, release ni licencia del repositorio. La contribución defendible es un gran diagnóstico sintético, chino primero, sobre seguimiento de preferencias explícitas e implícitas bajo distractores, con datos suficientes para auditorías posteriores y con trade-offs interesantes entre métodos. No demuestra personalidad infantil, lenguaje natural de niños, validez evolutiva, aprendizaje, bienestar, seguridad, beneficio para cuidadores, rendimiento con menores reales ni reproducción independiente de las figuras y resultados de entrenamiento.

Research question

Can LLMs infer and follow preferences of synthetic child personas, explicit or implicit, when the query appears after long irrelevant contexts, and do prompting with persona, LoRA, or a steering module improve performance?

Method

Chinese-English synthetic benchmark of preferences for profiles of 3-6 year olds. Qwen2.5-72B generates personas, preferences, queries, dialogues, and translations. WildChat provides real distractors. Five LLMs are evaluated with prompting, LoRA, and PSM using PC and four COE metrics judged by LLM; 600 cases receive a human check incompletely documented. The audit reviews two PDFs of 23 pages, a 2-page checklist, TeX, repository, code, and the 45,757 rows.

Sample: The paper rounds the benchmark to 46K cases and declares 29K personas. The released artifact contains 45,757 rows, 24,331 unique Chinese personas, and 14 subtopics. The human validation uses 600 examples without publishing labels or protocol.

Findings

  • Implicit preferences tend to be more difficult than explicit ones.
  • Irrelevant context tends to degrade PC, with non-monotonic curves.
  • Persona improves selected bars, but the effect depends on the model and dataset.
  • Interaction Scaffolding is the weakest COE dimension; DA is near the ceiling.
  • LoRA and PSM improve selected metrics, but change the error mix.
  • Published human-judge agreements are raw percentages without corrected reliability.
  • The published corpus has 828 length discrepancies and one role reversal.
  • The row-level splitter without seed expects to filter about 5,504 personas to both splits.
  • There are 28 milk-allergy/yogurt cases incompatible with an exhaustive safety review.
  • The metric extraction code fails and the LoRA/PSM training is not published.

Limitations

  • Personas, preferences, and language completely synthetic.
  • English automatically translated from Chinese.
  • Personality title without psychological construct or instrument.
  • Persona-preference relationship without alignment, conflict, or independence label.
  • Same Qwen generates data, translations, and main judge.
  • Human agreement without annotator count, instructions, kappa, prevalence, or uncertainty.
  • WildChat confounds memory, length, position, and distractor content.
  • WildChat IDs, filtering, and assembly absent.
  • 828 dialogues violate the declared number of rounds.
  • Split not grouped by persona and without seed.
  • Child risks and age-dependent safety out of scope.
  • Safety review without audit trail and medical failures present.
  • Ethics checklist contradictory with annotators and volunteer materials.
  • Point results without intervals, repetitions, or error bars.
  • LoRA/PSM hyperparameters absent.
  • Partial code, broken extractor, and no reproducible license.

What the study does not establish

  • Psychological personality of children or of the LLMs.
  • Real language, preferences, or behavior of 3-6 year old children.
  • Psychometric or developmental validity of PC/COE.
  • Independent human reliability of the judges.
  • Child safety or absence of medical and developmental harm.
  • Educational benefit, well-being, or caregiver support.
  • Performance in interaction with real minors.
  • Generalization outside the synthetic Chinese pipeline.
  • Memory effect separated from length, position, and noise.
  • Fine-tuning improvements free of persona leakage.
  • Independent reproduction of figures, training, or judge labels.

Traceability

Scope: Full text

Version: Findings of ACL 2026, pages 21282-21304; arXiv:2605.27805v1; final paper, Responsible NLP Checklist, complete arXiv source, repository commit 73e0c3e03b1a858b959949fd5c2feaff29a9c033 and all 45,757 released rows audited

Consulted source: https://aclanthology.org/2026.findings-acl.1070/

Review: Codex 48-page visual, Responsible NLP Checklist, complete TeX, 45,757-row data integrity, code, construct, judge, human evidence, ethics, safety and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Qwen2.5-72B-Instruct
  • Qwen2.5-3B-Instruct
  • Qwen3-4B-Instruct
  • LLaMA3.1-8B-Instruct
  • Mistral-7B-Instruct-v0.3
  • DeepSeek-R1-671B
  • BGE-large-zh-v1.5

Instruments and metrics

  • Preference Consistency
  • Emotional Adaptation
  • Interaction Scaffolding
  • Developmental Appropriateness
  • Engagement
  • Qwen2.5-72B LLM-as-a-judge
  • DeepSeek-R1 cross-judge check
  • FAISS semantic filtering
  • LoRA supervised fine-tuning
  • Persona Steer Model
  • Responsible NLP Checklist

Data used

  • ChildEval release: 45.757 bilingual rows
  • 24.331 unique Chinese persona strings
  • 45.755 English persona strings
  • 100 raw generation examples
  • WildChat-1M sampled contexts not released
  • 600 human-annotated cases not released

Evidence and location

  • Final publication, method, results, limitations, ethics, prompts, and figures: Findings of ACL 2026, 2026.findings-acl.1070, 23 pages, sha256 dc26777576f6cb49ee2d830e212a8d481bd1f6bfb9f7ed7cb6b086e79c6604a3
  • Preprint, TeX, and complete source: arXiv:2605.27805v1, 23 pages, source sha256 795344c0c847890797cdd733d6d827c8db43d773d19b0a3f981d7a2f7a9c6692
  • Ethics declarations and inconsistent responses about experiments and annotators: Responsible NLP Checklist, 2 pages, sha256 32665a9cb0ec8e43c693131a81d8ac14b690e677da54b9b80e5d74172ed25771
  • Dataset, scripts, prompts, integrity, safety, and reproducibility: github.com/ziyanluo/ChildEval commit 73e0c3e03b1a858b959949fd5c2feaff29a9c033, tree 0f0736d0caa95a6772f9d5828b59d055b46fa906
  • Complete independent audit: reports/verification/article-318-childeval-synthetic-child-construct-data-integrity-judge-human-ethics-safety-and-reproducibility-audit.json