PATS: Personality-Aware Teaching Strategies with Large Language Model Tutors

Applications, bias, and safety2026arXivApproved editorial review

Authors: Donya Rooein, Sankalan Pal Chowdhury, Mariia Eremeeva, Yuan Qin, Debora Nozza, Mrinmaya Sachan, Dirk Hovy

Keywords: Large Language Models, Personality, Persona, Model Evaluation

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

7
Authors
17
Findings
91
Limitations
13
Evidence

Editorial summary

English

PATS proposes that an LLM tutor should explicitly choose pedagogical strategies from a Big Five profile rather than merely answer questions. This review uses the definitive Findings of EACL 2026 publication, DOI 10.18653/v1/2026.findings-eacl.219, pages 4186–4211, and audits the official donya-rooein/PATS repository at commit 3924adf. The repository does release 3,840 simulated dialogues, 1,920 per task, 120 expert ratings, and a demo, but not the generation code, Prolific data, LLM evaluations, the 100-dialog author annotation sample, or a complete reproducible environment.

The real starting point is not a PATS intervention with students. The authors first examine private data from 110 participants aged 16–21, 281 GPT-4o tutor chats, and a 44-item Big Five questionnaire in a real course. They qualitatively conclude that the tutor mostly used question answering and that engagement was low, but provide no systematic coding procedure or engagement measure. In the appendix they attribute silence to low openness/high neuroticism and bullet-formatted language to the student having “summoned ChatGPT,” partly invoking AI detectors. These are speculative interpretations, not validated outcomes. PATS is subsequently evaluated only with simulated students, not those real learners.

The taxonomy links 17 strategies to each high or low OCEAN trait. It includes scaffolding, encouragement, direct instruction, role-play, friendly tone, choice, gamification, real-world relevance, progress tracking, and others. It draws on educational literature and filters techniques unsuitable for one-to-one tutoring. However, it turns continuous traits into ten binary stereotypes: low openness is described as “uncreative, inflexible, uninterested in learning,” and low agreeableness as “disinterested, inconsiderate, impolite, uncooperative.” High Agreeableness duplicates High Openness characteristics and recommends Collaborative Peer Teaching even though the filter claims to exclude group-dependent methods. There is no reproducible systematic review, expert panel, evidence grading per mapping, or causal validation that each strategy benefits the assigned profile.

The experiment uses two nominal third-grade tasks: describing ten Freepik images and discussing ten moral stories. For each task it generates all 32 binary Big Five combinations under three tutors, producing 640 dialogues per system and 1,920 per task. Gemini-2.0-Flash simulates the student in two stages: it generates five candidate replies and then selects the one most consistent with the profile. The same Gemini backend generates the tutors, including PATS and both baselines. PATS receives the true personality vector, chooses one trait, selects one or more permitted strategies, and creates a plan for the Responder. The personality baseline receives the same profile without the taxonomy; the simple baseline does not receive it. This tests privileged ground-truth conditioning, not reliable personality detection from real conversation, despite Figure 1 and parts of the text describing student “identification.”

Models and parameters are not fully pinned. Generation and evaluation ran from April to May 2025 using “GPT4o latest,” Llama 3.3 70B through Together, and Gemini-2.0-Flash through Google AI Studio with default hyperparameters. Exact snapshots, temperature, top-p, token limits, seeds, retries, and executable API code are not published. Gemini is selected as student simulator because GPT-4o classifies its traits better on average, creating LLM-judge dependence rather than validating child realism. The simulator uses explicit descriptors, theatrical body-language cues, and a second LLM call that chooses the most stereotypically profile-consistent response, making traits separable by construction.

The strongest descriptive evidence is that PATS produces different styles and a broader strategy range. Three authors review 100 dialogues and report 94.6% success at “mentioning and explaining” at least one selected strategy, with mean Cohen's kappa 0.610. That criterion verifies textual plan compliance, not learning, pedagogical appropriateness, or profile fidelity. Across the published PATS data, motivation and scaffolding dominate; only 12 of 17 strategies appear, High Agreeableness is never selected, and Low Openness appears once. Table 2 does not report literal frequencies: the JSON contains 786 selection events and 1,302 strategy mentions. Fractionally allocating each event across its k strategies yields 301.83 for motivation, 167.33 for scaffolding, and so forth; the paper truncates these to integers. Printed frequencies therefore total 783 rather than 786 without documenting the weighting rule.

Three LLM judges, GPT-4o, Llama 3.3 70B, and Gemini-2.0-Flash, compare PATS against each baseline in both orders; an order-dependent answer becomes a tie. For stories, Gemini and Llama favor PATS at roughly 60%, while GPT is close to a tie; image preferences are smaller, with several PATS bars around 37–48%. The caption claims significance from one-sided t-tests, but the repository contains neither the LLM choices nor the test code. The judge prompt explains that some tutors use targeted adaptive strategies, gives examples of their benefits, and discourages ties, potentially priming the measured preference. Gemini also evaluates Gemini-generated dialogues.

The initial human evaluation with 120 Prolific teachers is explicitly unreliable: only 42% pass the attention check and Fleiss kappa is 0.14–0.16. After exclusions, PATS receives 54.7% preference over the simple baseline and 54.3% over the personality baseline; role-play/friendly reaches p=0.002 over simple but p=0.060 over personality. Those data are not released. The evaluation supporting the conclusion uses only four expert teachers, all experienced with technology/AI. Each rates 30 comparisons, with 20 pairs per baseline and three votes per pair. Majorities favor PATS: against simple, 17/20 for motivation, 15/20 for personality fit, proactivity, and engagement, and 16/20 for empathy; against personality, 13/20 for motivation and fit, 12/20 for proactivity and engagement, and 15/20 for empathy. These are preferences over synthetic text, not student effects.

The JSON files reproduce those expert majorities, but the notebook is not executable end to end. It opens batch_Batch_*.json from the wrong directory, expects an LLMChoices key absent from all four released files, declares no environment, and retains outputs produced with different internal data. Its Fleiss function ignores its argument and captures a global i; it also converts ties into half-votes, creating impossible counts such as 2.5 annotators. Binomial tests treat three repeated ratings per dialogue as independent trials, score a tie as half a success, round the total, and do not correct ten comparisons. The section labeled Llama reuses human results_L1/results_L2 instead of results_L1_llama/results_L2_llama. Human chart proportions are recoverable, but the notebook's inference and LLM results should not be considered reproduced.

The defensible contribution is a design result: explicitly providing a taxonomy and separating strategy planning from response generation can produce dialogues that limited judges perceive as more motivating, empathetic, or adapted than less structured prompts. The study does not demonstrate improved learning, retention, real engagement, equity, or well-being. It also does not validate personality inference or safe psychological profiling of minors. PATS should be presented as a prompting prototype and synthetic corpus preferred in limited comparisons, not as an already validated effective personalized tutor.

Español

PATS propone que un tutor LLM no se limite a responder preguntas, sino que elija explícitamente estrategias pedagógicas en función de un perfil Big Five. Esta revisión usa la publicación definitiva en Findings of EACL 2026, DOI 10.18653/v1/2026.findings-eacl.219, páginas 4186–4211, y audita también el repositorio oficial donya-rooein/PATS en el commit 3924adf. El repositorio sí publica los 3.840 diálogos simulados, 1.920 por tarea, 120 valoraciones expertas y un demo, pero no el código de generación, los datos Prolific, las evaluaciones LLM, la muestra anotada de 100 diálogos ni un entorno reproducible completo.

El punto de partida real no es una intervención PATS con estudiantes. Los autores analizan primero datos privados de 110 participantes de 16–21 años, 281 chats de un tutor GPT-4o y un cuestionario Big Five de 44 ítems en un curso real. Concluyen cualitativamente que ese tutor usaba casi siempre preguntas y respuestas y que el engagement era bajo, pero no describen una codificación sistemática ni una métrica de engagement. En el apéndice atribuyen silencios a baja apertura/alto neuroticismo y lenguaje con viñetas a que el estudiante «summoned ChatGPT», apoyándose también en detectores de IA; son interpretaciones especulativas, no outcomes validados. PATS se evalúa después únicamente con estudiantes simulados, no con esos estudiantes reales.

La taxonomía enlaza 17 estrategias con cada rasgo OCEAN en nivel alto o bajo. Incluye scaffolding, motivación, instrucción directa, role-play, tono amistoso, elección, gamificación, relevancia real, seguimiento y otras. Su construcción parte de bibliografía educativa y filtra técnicas que no caben en tutoría individual. Sin embargo, la tabla convierte rasgos continuos en diez estereotipos binarios: por ejemplo, baja apertura se describe como «uncreative, inflexible, uninterested in learning» y baja amabilidad como «disinterested, inconsiderate, impolite, uncooperative». High Agreeableness copia características de High Openness y recomienda Collaborative Peer Teaching pese a que el propio filtro decía excluir estrategias dependientes de grupo. No hay revisión sistemática reproducible, panel de expertos, calidad de evidencia por mapping ni validación causal de que cada estrategia beneficie a ese perfil.

El experimento usa dos tareas para supuesto alumnado de tercer curso: describir diez imágenes de Freepik y comentar diez cuentos morales. Para cada tarea genera 32 combinaciones binarias de Big Five con tres tutores, 640 diálogos por sistema y 1.920 por tarea. Gemini-2.0-Flash simula al estudiante mediante dos pasos: genera cinco respuestas candidatas y luego selecciona la más coherente con el perfil. El mismo Gemini genera los tutores, incluidos PATS y sus dos baselines. PATS conoce el vector de personalidad verdadero, selecciona un rasgo, escoge una o varias estrategias permitidas y produce un plan que el Responder incorpora. El baseline de personalidad recibe el mismo perfil pero no la taxonomía; el simple no recibe el perfil. Por tanto se prueba generación condicionada con ground truth privilegiado, no detección fiable de personalidad a partir de una conversación real, aunque la Figura 1 y parte del texto hablen de «identificar» al estudiante.

Los modelos y parámetros tampoco están plenamente fijados. La generación y evaluación se hicieron de abril a mayo de 2025 con «GPT4o latest», Llama 3.3 70B en Together y Gemini-2.0-Flash en Google AI Studio, usando hiperparámetros por defecto. No se publican snapshots exactos, temperatura, top-p, max tokens, seeds, retries ni prompts/API completos como código ejecutable. Gemini se elige como simulador porque GPT-4o clasifica mejor sus rasgos en promedio, lo que introduce dependencia de un juez LLM y no valida realismo infantil. El simulador usa descriptores explícitos, body-language teatral y un segundo LLM que elige la respuesta más estereotípica; esto amplifica separabilidad por diseño.

La evidencia descriptiva más sólida es que PATS produce estilos diferentes y una gama más amplia de estrategias. Tres autores revisan 100 diálogos y reportan 94,6 % de éxito al «mencionar y explicar» al menos una estrategia seleccionada, con Cohen κ medio 0,610. Ese criterio verifica cumplimiento textual del plan, no aprendizaje, idoneidad pedagógica ni fidelidad al perfil. En los 640 diálogos PATS por tarea, motivación y scaffolding dominan; solo aparecen 12 de 17 estrategias, High Agreeableness nunca es seleccionado y Low Openness solo aparece una vez en los datos publicados. La Tabla 2 no contiene frecuencias literales: el JSON tiene 786 eventos de selección y 1.302 menciones de estrategia; al repartir cada evento entre sus k estrategias se obtienen frecuencias fraccionarias, 301,83 motivación, 167,33 scaffolding, etc., que el artículo trunca a enteros. Así las columnas impresas suman 783, no 786, sin explicar la ponderación.

Tres LLM, GPT-4o, Llama 3.3 70B y Gemini-2.0-Flash, comparan PATS con cada baseline en orden invertido; si cambia la respuesta se marca empate. En historias, Gemini y Llama favorecen PATS alrededor del 60 %, mientras GPT queda cerca de empate; en imágenes las preferencias PATS son menores y varias barras rondan 37–48 %. El caption afirma significancia mediante t-tests unilaterales, pero el repositorio no contiene las elecciones LLM ni el código de esos tests. El prompt del juez explica que ciertos tutores incorporan estrategias adaptadas, da ejemplos de sus beneficios y pide evitar empates, lo que puede inducir la preferencia que mide. Gemini además evalúa diálogos generados por Gemini.

La primera evaluación humana con 120 docentes de Prolific es explícitamente poco fiable: solo 42 % supera el control de atención y Fleiss κ es 0,14–0,16. Tras excluir fallos, PATS logra 54,7 % frente al baseline simple y 54,3 % frente al de personalidad; role-play/friendly llega a p=0,002 frente al simple pero p=0,060 frente al baseline de personalidad. Estos datos no se publican. La evaluación que sustenta la conclusión usa solo cuatro docentes expertos, todos con experiencia tecnológica/IA. Cada uno valora 30 comparaciones; hay 20 pares por baseline y tres votos por par. La mayoría favorece PATS: frente al simple, 17/20 en motivación, 15/20 en ajuste, proactividad y engagement, y 16/20 en empatía; frente al baseline de personalidad, 13/20 en motivación y ajuste, 12/20 en proactividad y engagement, y 15/20 en empatía. Es preferencia sobre textos sintéticos, no efecto en estudiantes.

Los JSON permiten reproducir esas mayorías expertas, pero el notebook no es ejecutable de extremo a extremo. Abre batch_Batch_*.json desde el directorio equivocado, espera una clave LLMChoices ausente en los cuatro archivos, no declara dependencias y conserva outputs generados con datos internos distintos. La función Fleiss ignora su parámetro y depende de la variable global i; además convierte empates en medios votos, creando conteos imposibles como 2,5 anotadores. Los binomiales cuentan tres votos repetidos por diálogo como ensayos independientes, convierten empate en medio éxito, redondean y no corrigen diez comparaciones. La sección rotulada Llama vuelve a usar results_L1/results_L2 humanos en lugar de results_L1_llama/results_L2_llama. Por ello las barras humanas son recuperables, pero no deben aceptarse las pruebas inferenciales del notebook ni las cifras LLM como reproducidas.

La aportación defendible es de diseño: hacer explícita una taxonomía y separar planificación de respuesta puede generar diálogos que jueces perciben como más motivadores, empáticos o adaptados que prompts menos estructurados. No se demuestra mejora de aprendizaje, retención, engagement real, equidad o bienestar. Tampoco se valida inferencia de personalidad ni uso seguro de perfiles psicológicos en menores. PATS debe presentarse como prototipo de prompting y corpus sintético preferido en comparaciones limitadas, no como tutor personalizado eficaz ya validado.

Research question

Can an LLM tutor that knows a Big Five profile and explicitly selects pedagogical strategies produce simulated dialogues perceived as better adapted than a simple prompt or a tutor that knows the profile but does not receive a taxonomy?

Method

A taxonomy of 17 strategies is derived for ten high/low Big Five states. Gemini-2.0-Flash simulates third-grade students and generates three versions of each dialogue, simple, personality prompt, and PATS, across ten images and ten stories, for 32 binary profiles and 3,840 dialogues. Three LLMs evaluate pairwise preferences with inverted order. Three authors annotate 100 dialogues for strategy compliance; 120 Prolific teachers conduct an unreliable preliminary evaluation; four expert teachers rate 40 pairs across five dimensions.

Sample: The factorial generation covers two tasks × ten topics × 32 profiles × three tutors = 3,840 dialogues. The implementation analysis uses 100 randomly chosen dialogues without a seed. The Prolific study recruits 120 teachers, but only 42 % passes the control. The expert study includes four teachers and 40 pairs; each teacher rates 30 and each pair receives three votes. There are no real students exposed to PATS or learning measures.

Findings

  • The repository contains exactly 320 dialogues for each task × system combination, covering the 32 profiles ten times.
  • Three authors report 94.6 % textual compliance with at least one planned strategy in 100 dialogues, with mean Cohen κ 0.610.
  • Motivational Encouragement and Step-by-Step Scaffolding dominate PATS selection.
  • Only 12 of 17 strategies appear in the published dialogues.
  • High Agreeableness is never selected by the Strategizer and Low Openness appears only once.
  • Table 2 represents 786 events through fractional weighting by number of strategies and truncation; the printed integers sum to 783.
  • In stories, Gemini and Llama favor PATS around 60 %, while GPT-4o remains close to a tie.
  • In images, LLM preferences are weaker and several PATS rates are between 37 % and 48 %.
  • Agreement among LLM judges is reported at Fleiss κ 0.35–0.41 using inverted runs of the same model as if they were separate evaluators.
  • In Prolific, only 42 % passes attention and Fleiss κ is 0.14–0.16.
  • Aggregated Prolific preference is 54.7 % against simple and 54.3 % against the personality baseline.
  • Role-play/Friendly is significant against simple (p=0.002) but not against the personality baseline (p=0.060).
  • Against simple, experts favor PATS in 17/20 pairs for motivation and 15–16/20 in the other dimensions.
  • Against the personality baseline, experts favor PATS in 12–15/20 pairs depending on dimension.
  • The majority expert percentages are reproduced from the four published JSONs.
  • The notebook cannot reproduce LLM preferences or their tests with the published data.
  • PATS changes the shape of the dialogues and increases explicit strategies, but the study does not measure learning.

Limitations

  • The main evaluation of PATS does not include real students.
  • The 110 students and 281 real chats only motivate the problem and do not receive the intervention.
  • No quantitative metric of engagement for the real classroom is defined.
  • The claim of low participation comes from qualitative analysis without a coding protocol.
  • No coders, rubric, disagreements, or reliability are reported for the 281 chats.
  • Attributing silences to openness/neuroticism is speculative and may pathologize ambiguous behavior.
  • Inferring ChatGPT use from bullet points, elaborate language, and AI detectors is not reliable verification.
  • The real data are private and cannot be audited.
  • The paper says high if >3, while Figure 5 says high if ≥3 and low if ≤3.
  • A score of exactly 3 is left unclassified under one rule and belongs to both groups under the other.
  • Dichotomizing Big Five loses information and may create artificially distinct profiles.
  • The ten most common profiles are used for human evaluation, introducing selection by prevalence.
  • The taxonomy is not derived through a reproducible systematic review.
  • No search terms, databases, screening, dates, or inclusion diagram are published.
  • There is no quality or strength-of-evidence evaluation by mapping.
  • No independent panel of teachers or psychometricians was consulted to validate the taxonomy.
  • The cited bibliography relates traits, preferences, and outcomes, but does not always prove that the assigned strategy causes benefit.
  • Low trait descriptors are pejorative and stereotypical.
  • Low Openness is equated with lack of creativity, inflexibility, and disinterest.
  • Low Agreeableness is equated with inconsiderateness, discourtesy, and lack of cooperation.
  • High Agreeableness repeats characteristics of High Openness in Table 1.
  • Collaborative Peer Teaching is retained even though the filter declares it excludes group strategies.
  • The taxonomy does not model interaction among five traits, context, culture, language, age, ability, or disability.
  • The Strategizer acts on a single trait at a time and may ignore tensions of the full profile.
  • PATS receives the true profile; it does not have to infer it from conversation.
  • The supposed personality detection shown in Figure 1 is not evaluated as part of the PATS pipeline.
  • Applying the real system would require a questionnaire or psychological inference with privacy, consent, and error risks.
  • The simulator knows explicit personality descriptions and is instructed to portray them.
  • Each student turn is selected from five candidates for coherence with the profile, amplifying stereotypes.
  • Response selection does not evaluate realism of age, knowledge, pedagogy, or diversity, only persona consistency.
  • GPT-4o judges the personality of outputs that are used to choose Gemini as the simulator, creating judge dependence.
  • F1 for trait identification does not equal behavioral or human fidelity.
  • There is no comparison of the simulator with real children's responses.
  • The same Gemini generates the student, PATS, and baselines, creating shared style and errors.
  • Gemini also acts as a judge over dialogues produced by Gemini.
  • The models are identified as GPT4o latest, Gemini-2.0-Flash, and Llama 3.3 70B without complete snapshots.
  • Default hyperparameters are used, which may change by provider and date.
  • No temperature, top-p, max tokens, seed, safety settings, retries, or failed responses are reported.
  • There is no generation code or API configuration in the repository.
  • There is no requirements.txt, lockfile, pyproject, or executable environment instructions.
  • The evaluation is only in English and in two narrow tasks for third grade.
  • The images are AI-generated and are hotlinked from Freepik without a snapshot or per-element license metadata.
  • The stories come from an external site and cultural representativeness or detailed licensing is not discussed.
  • There is no control for harmful content, cultural biases, or adequacy of the stories/images.
  • The 94.6 % criterion only requires that the output mentions and explains some selected strategy.
  • Plan compliance does not measure quality, adequacy, learning, or benefit for the profile.
  • The three compliance annotators are authors, without independent evaluation.
  • The random sample of 100 has no seed and no published IDs.
  • The labels of that sample are not in the repository.
  • The mean Cohen κ=0.610 is not broken down by strategy, task, or annotator pair.
  • PATS only uses 12 of 17 strategies, showing incomplete coverage.
  • High Agreeableness is never selected and Low Openness almost never, despite evaluating all 32 profiles in a balanced way.
  • Table 2 calls frequencies what are truncated fractional assignments without documenting the formula.
  • The integers in Table 2 sum to 783 although the JSON contains 786 selection events.
  • The LLM evaluation prompt explains the expected advantages of personalized strategies and may prime PATS.
  • LLM judges are asked to avoid Both are Equal, biasing against ties.
  • Treating the change when inverting order as a tie preserves order sensitivity without modeling it.
  • The same model with two orders is treated as separate evaluators for Fleiss κ, even though they are not independent.
  • Figure 3 uses one-tailed t-tests over categorical preferences without justifying unit, normality, or independence.
  • No data or code for the LLM tests are published.
  • Multiple comparisons by model, task, and baseline are not clearly corrected.
  • The Prolific study fails its own control: only 42 % passes attention.
  • Fleiss κ 0.14–0.16 shows very low agreement in Prolific.
  • Excluding attention failures after the result may change the evaluated population.
  • Prolific data are not published to verify exclusions, payments, or tests.
  • Four experts are a very small sample and all declare Tech/AI experience.
  • Each expert evaluates 30 pairs, so their repeated votes are not independent.
  • Each of the 40 pairs receives three votes and the five dimensions share the same text and evaluators.
  • Majority percentages are based only on 20 pairs per baseline.
  • Teacher, dialogue, task, or personality effects are not modeled in a multilevel analysis.
  • The binomial tests in the notebook treat repeated votes as 60 independent trials per dimension.
  • The notebook assigns half a success to a tie and rounds before the binomial test.
  • It does not correct ten human tests of dimension × baseline.
  • The notebook's Fleiss function ignores the index parameter and depends on a global variable i.
  • The Fleiss calculation converts ties into half votes and produces non-integer annotator counts.
  • The notebook looks for batch_Batch_i.json outside Data/Survey if run from its location.
  • The published JSONs do not contain the LLMChoices key that several cells require.
  • The notebook retains outputs produced with an internal data version different from the published one.
  • The binomial section for Llama mistakenly uses the human results_L1/results_L2 dictionaries.
  • There are no automated tests, analytical CI, or JSON schema validation.
  • The README claims to include code/scripts and data to reproduce figures, but the actual scope is partial.
  • The README citation remains arXiv and not the definitive EACL publication.
  • The root MIT license does not clarify the coexistence with Gemini dialogues declared CC BY 4.0 and external assets.
  • Content accuracy, learning, pre/post, retention, transfer, or task time are not measured.
  • Real behavioral engagement, dropout, student satisfaction, or cognitive load are not measured.
  • Text preference by a judge does not equal pedagogical efficacy.
  • Longer, more proactive, or more explicit responses may win preferences even if they do not improve learning.
  • Adverse effects of labeling or treating a student according to a possibly incorrect profile are not evaluated.
  • IRB details are withheld; approval is declared, but protocol, consent, or assent cannot be verified.
  • The real sample possibly includes minors, but parental consent and age-based handling are not detailed.
  • Generalization to other languages, ages, cultures, subjects, special needs, or classrooms remains untested.

What the study does not establish

  • It does not establish that adapting strategies to Big Five improves learning or retention.
  • It does not demonstrate greater real engagement in students.
  • It does not validate PATS through a controlled classroom intervention.
  • It does not demonstrate that its taxonomy is correct, complete, or causal.
  • It does not validate that low traits imply the stereotypical behaviors used in prompts.
  • It does not demonstrate that an LLM can reliably infer the personality of a real student.
  • It does not demonstrate safety, equity, or benefit of psychological profiling of minors.
  • It does not establish that Gemini outputs faithfully simulate third-grade students.
  • It does not demonstrate that textual compliance with a strategy is effective pedagogical implementation.
  • It does not demonstrate that role-play is of high impact on learning; it only obtains limited preference in certain contrasts.
  • It does not establish reproducible significance of LLM evaluations with the public artifacts.
  • It does not convert the Prolific study into reliable evidence after 42 % attention and low κ.
  • It does not generalize from four experts and 40 pairs to teachers or students as a whole.
  • It does not allow reproducing the full pipeline, costs, models, and statistics from the repository.
  • It does not justify deploying PATS as an autonomous tutor or teacher replacement.

Traceability

Scope: Full text

Version: Findings of EACL 2026, DOI 10.18653/v1/2026.findings-eacl.219, pp. 4186–4211; 26 pages

Consulted source: https://aclanthology.org/2026.findings-eacl.219.pdf

Review: Codex full-text, bilingual-fidelity, definitive-ACL, 26-page visual, taxonomy-evidence, simulator-validity, classroom-claim, human-sampling, LLM-judge, statistical-independence, repository-code, notebook, JSON-coverage, frequency-recomputation, licensing and artifact-reproducibility audit; data validation and quality controls applied, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Google Gemini-2.0-Flash via Google AI Studio, exact dated snapshot not reported, used for all student and tutor dialogue generation
  • OpenAI GPT-4o latest via API, queried April–May 2025, exact snapshot not reported, used as personality classifier and dialogue judge
  • Meta Llama 3.3 70B via together.ai, exact provider snapshot not reported, used as dialogue judge
  • Gemini-2.0-Flash used again as dialogue judge over Gemini-generated conversations

Instruments and metrics

  • 44-item Big Five questionnaire scored on a 1–5 scale in the private classroom dataset
  • Binary high/low OCEAN profile construction with an internally inconsistent threshold at score 3
  • Author-built taxonomy mapping 17 pedagogical strategies to ten high/low trait states
  • Two-stage Gemini student simulator: five candidate utterances followed by profile-consistency selection
  • PATS Strategizer: trait selection, strategy selection, and implementation planning
  • PATS Responder with last four utterances plus summary of earlier context
  • Simple-prompt and personality-aware prompt baselines
  • Image-description and moral-story tasks for a nominal third-grade learner
  • GPT-4o personality classification F1 used to select the simulator backend
  • Three-author binary assessment of whether any planned strategy is mentioned and explained
  • Pairwise LLM preference with reversed order and order disagreement converted to tie
  • Five expert dimensions: motivation/support, personality suitability, proactivity, engagement, and empathy
  • Prolific attention checks and expert-teacher majority voting

Data used

  • Private classroom data: 110 participants aged 16–21, Big Five surveys, and 281 anonymized GPT-4o tutor chats
  • Ten Freepik image subjects and ten children’s moral-story subjects
  • 3,840 released synthetic dialogues: 320 profiles/subjects × 3 systems × 2 tasks
  • 640 PATS dialogues per task with strategy-plan records
  • Author annotation sample of 100 dialogues, labels not released
  • Preliminary Prolific study with 120 teachers and 200 dialogue comparisons, raw data not released
  • Expert study with four teachers, 40 unique pairs, three ratings per pair, and 120 released comparison records
  • Official repository donya-rooein/PATS at commit 3924adf32ab21ae79e027ffc6711798d56c3d0ba

Evidence and location

  • Definitive publication, DOI and pages: ACL Anthology 2026.findings-eacl.219, DOI 10.18653/v1/2026.findings-eacl.219, pp. 4186–4211
  • Real classroom data and Big Five threshold: Findings EACL paper, Section 2 and Appendix B–C
  • Taxonomy and mappings: Findings EACL paper, Section 3, Tables 1 and 6
  • Student model, PATS and baselines: Findings EACL paper, Sections 4.1–4.3 and Appendix E
  • Tasks, 32 profiles and 3,840 dialogues: Findings EACL paper, Section 4.4; official repository Data/Tasks
  • Models, dates and default parameters: Findings EACL paper, Appendix A
  • Strategy compliance and distribution: Findings EACL paper, Section 5.1, Tables 2 and Figures 2; repository Data/Tasks/*_L3.json
  • Implicit weighting of Table 2: Independent recomputation from 786 strategy-selection records in official repository commit 3924adf; 1/k allocation reproduces the printed frequencies before truncation
  • LLM evaluation and judge prompts: Findings EACL paper, Sections 4.5 and 5.2, Figure 3, Appendix F.1.4 and F.2.2
  • Prolific and low reliability: Findings EACL paper, Section 5.2 and Appendix F.1
  • Four experts and majorities: Findings EACL paper, Section 5.2, Figure 4, Table 7; repository Data/Survey/batch_Batch_1–4.json
  • Reproducibility failures and notebook code: Official repository commit 3924adf, Analysis/Analysis Expert Annotation.ipynb and released Data/Survey schema, audited 15 July 2026
  • Declared limitations and ethics: Findings EACL paper, Limitations and Ethical Considerations sections