Automatic Scoring of an Open-Response Measure of Advanced Mind-Reading Using Large Language Models

Evaluation and psychometric validity2025ACL AnthologyApproved editorial review

Authors: Yixiao Wang, Russel Dsouza, Robert Lee, Ian Apperly, Rory T. Devine, Sanne W. van der Kleij, Mark Lee

Keywords: Large Language Models, Theory of Mind, Psychometrics, Automated Scoring, Open-Response Assessment, LLM Evaluation

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

7
Authors
13
Findings
23
Limitations
11
Evidence

Editorial summary

English

This CLPsych 2025 paper does not study personality or theory of mind inside LLMs; it uses models as binary graders of human open-ended answers to ten social narratives intended to measure advanced mind-reading. A total of 1,733 people aged 13-30 participated through schools or Prolific. One lead coder and four trained coders assigned labels; before independent coding they reached kappa greater than .7 against the lead, who resolved discrepancies. Each prompt includes the narrative, question, detailed rubric, and answer, and the model returns 0/1. The paper compares plain/XML/JSON, three rubric versions, zero-shot, 10-shot, 50-shot, and fine-tuning. It reports 89.4% macro accuracy for zero-shot GPT-4o and 92.8% after fine-tuning, with kappa .83. RoBERTa-large reaches 92.0%, GPT-4o-mini 90.5%, Llama-3.2-3B 91.1%, Mistral 88.7%, Phi-3.5 83.8%, Phi-4 87.5%, and Longformer 86.7%. GPT-4o paraphrase augmentation helps Longformer, Mistral, and Phi-3.5, barely changes Phi-4, and slightly harms Llama. These are promising fixed-rubric imitation results, but artifact audit blocks the paper's broader interpretation. Code creates 17,330 response opportunities, randomly removes 3,078 majority-class rows before splitting, and publishes 14,252 balanced rows. The actual split is 11,543/1,283/1,426-80.99%/9.00%/10.01%, not 80/10/10. The pipeline also drops participant IDs before `train_test_split`; because each person answers all ten narratives, participant-specific style can occur in train and test. All ten items and rubrics appear in every split, so the experiment tests new answers to known items rather than transfer to new items or instruments. Three exact prompts overlap train/validation and one overlaps train/test. Although test is globally 50/50, its per-item majority baselines average 62.03% and range from 50.32% to 77.70%; zero-shot Llama at 64.3% is only 2.27 points higher. The paper calls several differences significant without tests, intervals, replications, or participant/item bootstrap. Its human comparison is more problematic: kappa is computed only on cases where GPT-4o and the coder disagreed, conditioning the sample, with the lead coder as reference. The prose says GPT-4o wins except on item 5, but Table 10 also has the human ahead on item 3 and contains no item 7 despite listing it in the text. The repository also cannot reproduce the headline: it lacks the OpenAI pipeline, predictions, results, few-shot/format and BERT/RoBERTa code; the paraphrase dataset returns 401, `requirements.txt` is missing, scripts contain broken paths, README swaps Phi/Mistral links, and checkpoints are incomplete. The defensible conclusion is that fine-tuned models can approximate a lead coder on ten known questions in a reduced balanced corpus. The work does not validate the test, generalization to new participants, items, demographic groups, or clinics, or use as a screening tool.

Español

Este trabajo de CLPsych 2025 no estudia personalidad ni teoría de la mente dentro de los LLM: usa modelos como calificadores binarios de respuestas humanas abiertas a diez narrativas sociales que pretenden medir mind-reading avanzado. Participaron 1.733 personas de 13 a 30 años, en centros educativos o Prolific. Un lead coder y cuatro codificadores entrenados asignaron etiquetas; antes de trabajar independientemente alcanzaron κ>0,7 frente al lead y este resolvió discrepancias. Cada prompt contiene narrativa, pregunta, rúbrica detallada y respuesta; el modelo devuelve 0/1. El paper compara plain/XML/JSON, tres versiones de rúbrica, zero-shot, 10-shot, 50-shot y fine-tuning. Declara 89,4 % macro de accuracy para GPT-4o zero-shot y 92,8 % tras fine-tuning, con κ=0,83. RoBERTa-large alcanza 92,0 %, GPT-4o-mini 90,5 %, Llama-3.2-3B 91,1 %, Mistral 88,7 %, Phi-3.5 83,8 %, Phi-4 87,5 % y Longformer 86,7 %. La augmentación mediante paráfrasis GPT-4o ayuda a Longformer, Mistral y Phi-3.5, apenas cambia Phi-4 y perjudica ligeramente a Llama. Son resultados prometedores para imitar una rúbrica fija, pero la auditoría del artefacto impide la interpretación amplia del artículo. El código genera 17.330 oportunidades de respuesta, elimina aleatoriamente 3.078 de la clase mayoritaria antes de dividir y publica 14.252 filas equilibradas. La partición real es 11.543/1.283/1.426, 80,99 %/9,00 %/10,01 %, no 80/10/10. Además, el pipeline descarta los IDs de participante antes de `train_test_split`; como cada persona responde las diez narrativas, sus estilos pueden aparecer en train y test. Los diez ítems y sus rúbricas están en los tres splits, por lo que el experimento mide nuevas respuestas a ítems conocidos, no transferencia a nuevos ítems o instrumentos. El dataset contiene tres prompts idénticos entre train/validation y uno entre train/test. Aunque test es 50/50 global, sus baselines mayoritarios por ítem promedian 62,03 % y varían de 50,32 % a 77,70 %; Llama zero-shot 64,3 % queda solo 2,27 puntos por encima. El paper llama significativas varias diferencias sin tests, intervalos, repeticiones o bootstrap por participante e ítem. La comparación humana es todavía más problemática: calcula kappa solo en casos donde GPT-4o y el codificador discrepaban, condicionando la muestra, y usa al lead coder como referencia. El texto dice que GPT-4o gana salvo en el ítem 5, pero Table 10 muestra que el humano también gana en el 3 y no contiene el ítem 7 que la prosa enumera. El repositorio tampoco reproduce el headline: carece de pipeline OpenAI, predicciones, resultados, código few-shot/formato y BERT/RoBERTa; el dataset de paráfrasis devuelve 401, falta `requirements.txt`, los scripts tienen rutas rotas, README intercambia enlaces Phi/Mistral y hay checkpoints incompletos. La conclusión defendible es que modelos ajustados pueden aproximar el criterio de un lead coder para diez preguntas conocidas en un corpus reducido y balanceado. No se valida el test, la generalización a participantes nuevos, ítems nuevos, grupos demográficos o clínicas, ni el uso como herramienta de cribado.

Research question

With what accuracy and agreement with a lead coder can various models score human open-ended responses on ten mind-reading items, and how do results change with format, examples, fine-tuning, and synthetic paraphrases?

Method

Ten social narratives are administered to 1,733 participants aged 13-30. A lead coder and four trained coders assign binary labels. The artifact concatenates story, question, rubric, and response, randomly removes examples from the global majority class, splits individual rows with seed 42, and evaluates macro accuracy per item. GPT-4o/mini prompts, open models with LoRA, ten per-item BERT/RoBERTa classifiers, and GPT-4o paraphrases added only to train are tested. The audit reproduces the profile of the public dataset, examines split/duplicates/baselines, and contrasts code, models, and availability.

Sample: 1,733 participants aged 13-30, recruited in schools or Prolific, with ten potential responses each. No exact distribution of age, country, culture, gender, ethnicity, diagnosis, site, or compensation is published. The released dataset retains 14,252 responses after undersampling: 11,543 train, 1,283 validation, and 1,426 test; it omits IDs and demographics.

Findings

  • The authoritative source is CLPsych 2025, DOI 10.18653/v1/2025.clpsych-1.7, pages 79-89; the 11 pages were rendered and visually inspected.
  • GPT-4o zero-shot achieves 89.4% macro; XML 0.84 versus plain 0.82 and JSON 0.83, and the original rubric 0.88 versus 0.86/0.85.
  • Ten shots barely change GPT-4o 89.4 to 89.5 and raise GPT-4o-mini 79.7 to 81.4; 50 shots reduce both to 88.17 and 80.1.
  • Fine-tuned GPT-4o achieves 92.8%; RoBERTa-large 92.0%, Llama 91.1%, GPT-4o-mini 90.5%, BERT/RoBERTa 90.2-92.0%.
  • Augmentation changes Longformer 86.7 to 91.6, Mistral 88.7 to 91.6, Phi-3.5 83.8 to 90.1, Llama 91.1 to 90.5, and Phi-4 87.5 to 87.6.
  • The public artifact removes 3,078 of 17,330 rows by global undersampling and retains 14,252, exactly balanced per split.
  • The real split is 80.99%/9.00%/10.01%, not 80/10/10.
  • The pipeline removes participant ID before splitting individual responses and does not implement group split.
  • All splits contain the same ten stories, questions, and rubrics; there is no item-held-out evaluation.
  • There are three exact prompts repeated train/validation and one train/test.
  • The majority baseline per item on test averages 62.03%, despite the 50/50 global balance.
  • Table 10 is calculated only on GPT/human disagreements; its prose omits that the human wins item 3 and mentions an absent item 7.
  • The scripts depend on a nonexistent requirements.txt and a nonexistent config path; central pipelines and outputs are missing.

Limitations

  • Splitting by response does not prevent one person from appearing in several splits.
  • Without public IDs, participant overlap cannot be measured or corrected.
  • Fine-tuning and test share the same ten items; generalization to new questions is not tested.
  • Four identical prompts cross splits, including one train/test.
  • Undersampling removes 17.76% of responses and alters prevalence and operating metrics.
  • Macro accuracy does not inform sensitivity, specificity, calibration, false positives/negatives, or clinical costs.
  • Per-item baselines reach 77.70% and are not compared in the paper.
  • There are no tests, intervals, repetitions, bootstrap per participant/item, or multiple correction for significance.
  • Format and rubric decisions do not describe a separate holdout; there is a risk of selection over evaluation.
  • The paper does not clarify whether the few-shot examples come only from train.
  • Kappa on cases selected by disagreement is not global reliability nor comparable with the initial human kappa.
  • Lead-coder agreement measures fidelity to the referent, not independent psychological truth.
  • Table 10 partially contradicts the prose and only covers two coders and nine coder-item combinations.
  • Dimensionality, internal consistency, test-retest, convergent/discriminant validity, invariance, or clinical validity are not evaluated.
  • No analyses by age, site, culture, gender, ethnicity, language, or diagnosis are published.
  • Ten narratives and a 13-30 sample do not support universal generalization.
  • A single coder validates 500 paraphrases with raw agreement >90%; kappa, intervals, and audit of the remainder are missing.
  • GPT-4o generates paraphrases and is also the scorer, a possible provider-style advantage.
  • The paraphrase dataset is not currently public; the base has no dataset card or declared license.
  • The repository lacks requirements.txt, outputs, and several implementations; README and checkpoints have errors.
  • The public model cards are empty templates without intended use, data, risks, or clinical limits.
  • Authentic narratives of youth and responses from age 13 are published, but the paper only indicates ethical approval and does not detail publication consent, anonymization, withdrawal, or governance.
  • The ID ranges in Table 10 reach 2,157 although 1,733 participants are reported; the numbering is not explained.

What the study does not establish

  • It does not demonstrate that the LLM has theory of mind or understands mental states as a person does.
  • It does not study or demonstrate LLM personality.
  • It does not psychometrically validate the ten-item instrument.
  • It does not demonstrate generalization to new participants due to the split without groups.
  • It does not demonstrate transfer to new items, tests, languages, or domains.
  • It does not establish general superiority over human coders with the disagreement-selected kappa.
  • It does not test the significance of accuracy differences or augmentation effects.
  • It does not establish safety or utility as a clinical, diagnostic, or screening tool.
  • It does not demonstrate equity across ages, demographics, or clinical conditions.
  • It does not reproduce end-to-end the 92.8% of GPT-4o from the public artifact.

Traceability

Scope: Full text

Version: Proceedings of CLPsych 2025, Anthology ID 2025.clpsych-1.7, DOI 10.18653/v1/2025.clpsych-1.7, pages 79-89, 11 pages

Consulted source: https://aclanthology.org/2025.clpsych-1.7/

Review: Codex complete bilingual full-text fidelity pass using the published CLPsych version, all-page visual inspection, official Git history audit, Hugging Face split profiling, exact label and item-baseline analysis, cross-split duplicate inspection, participant-group leakage review, model-repository inventory, reproducibility audit, kappa-table reconciliation, psychometric-validity assessment, ethical-reporting review, and scope-boundary correction; summaries written from the paper and artifact evidence rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • openai/gpt-4o-2024-08-06
  • openai/gpt-4o-mini-2024-07-18
  • allenai/longformer-base-4096
  • Paper label meta/llama-3.2-3B-instruct; released fine-tuned artifact uses NousResearch/Hermes-3-Llama-3.2-3B
  • mistralai/Mistral-7B-Instruct-v0.3
  • microsoft/Phi-3.5-mini-instruct
  • microsoft/phi-4
  • BERT-base and BERT-large, ten item-specific classifiers each
  • RoBERTa-base and RoBERTa-large, ten item-specific classifiers each

Instruments and metrics

  • Ten open-response advanced mind-reading social-narrative items
  • Item-specific binary coding manuals
  • Human coding with lead-coder adjudication and initial Cohen's kappa greater than .7
  • Plain, XML, and JSON prompt formats
  • Original, rephrased, and GPT-4o-summarized grading schemes
  • Zero-shot, 10-shot, and 50-shot prompting
  • LoRA and OpenAI fine-tuning
  • GPT-4o training-response paraphrase augmentation
  • Accuracy and selected-subset Cohen's kappa
  • Independent response-split, duplicate, prevalence, and majority-baseline audit

Data used

  • Reported raw design: 1,733 participants x 10 items = 17,330 response opportunities
  • Hugging Face rshwndsz/ToM-auto-scoring-base commit 6c2334d2854cd1e50b5cb1565524079e780e5fc2: 14,252 balanced derived rows
  • Train 11,543; validation 1,283; test 1,426
  • Hugging Face rshwndsz/ToM-auto-scoring-paraphrased, inaccessible with HTTP 401 on 15 July 2026
  • Public fine-tuned model repositories, several incomplete or mismatched with README labels

Evidence and location

  • Abstract, scope, sample, and research questions: Published pages 79-80, Abstract and Introduction
  • Data, coding, and per-item difficulty: Published pages 81-82, Section 3.1 and Tables 1-2
  • Models, prompts, declared split, fine-tuning, and paraphrases: Published pages 83-84, Sections 3.2-3.5
  • Zero/few-shot, fine-tuning, and augmentation results: Published pages 84-86, Tables 4-9
  • Human comparison, selected kappa, and prose error: Published pages 85-87, Section 4.4 and Table 10
  • Acknowledged limitation and ethical approval: Published pages 87-88, Sections 7-8
  • Undersampling, row split, loss of IDs, real split, and absence of item holdout: Official repository commits 601612c and 69152da, odysseus/data.py audited 15 July 2026
  • 14,252 rows, balance, per-item baselines, and cross-split duplicates: Hugging Face base dataset commit 6c2334d2854cd1e50b5cb1565524079e780e5fc2 independently profiled 15 July 2026
  • Inaccessible, incomplete, broken, or mislinked artifacts: Official GitHub and ten linked Hugging Face model repositories inventoried 15 July 2026
  • Comprehensive validity and reproduction audit: reports/verification/article-192-psychometric-split-and-artifact-audit.json
  • Complete visual inspection: All 11 published PDF pages rendered and visually inspected on 15 July 2026