Quantifying Data Contamination in Psychometric Evaluations of LLMs

Evaluation and psychometric validity2026arXivApproved editorial review

Authors: Jongwook Han, Woojung Song, Jonggeun Lee, Yohan Jo

Keywords: data contamination, psychometric evaluation, item memorization, scoring knowledge, target-score matching, LLM evaluation

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

4
Authors
4
Findings
4
Limitations
3
Evidence

Editorial summary

English

The paper proposes measuring alleged contamination in LLM psychometric evaluation through three task families: semantic or keyword reconstruction of inventory items, recovery of item-to-dimension and option-scoring knowledge, and response adjustment to match a target score. It evaluates 21 models on the BFI-44, PVQ-40, MFQ, and Short Dark Triad at temperature 0.7, with three runs and 95% confidence intervals. Reported overall averages are 0.31 for semantic reconstruction, 0.39 for keyword recovery, 0.94 F1 for item–dimension association, 0.44 MAE for option-score recovery, and 0.45 MAE for target matching; some advanced models reach approximately 0.1–0.2 target MAE. BFI and PVQ generally yield stronger results than MFQ and SD-3. These tasks, however, expose item text, dimensions, response options, or targets, so semantic inference and reasoning can solve them without training-set exposure. The paper’s own generic dimension-description baseline scores 0.32 on semantic reconstruction, slightly above the 0.31 mean interpreted as memorization. The design provides no training-corpus access, canaries, matched newly written inventories, or causal link between task scores and bias in downstream evaluations. It therefore supplies a useful diagnostic of instrument familiarity and functional scoring knowledge, but does not by itself identify contamination or prove memorization. The official repository contains runnable code and 19 tests that pass in a clean environment, but it does not commit the paper’s result outputs, and full reproduction depends on live, mutable model APIs.

Español

El trabajo propone medir una supuesta contaminación de evaluaciones psicométricas de LLM mediante tres familias de tareas: reconstrucción semántica o de palabras en ítems, recuperación de la relación entre ítems y dimensiones o puntuaciones, y ajuste de respuestas para alcanzar una puntuación objetivo. Evalúa 21 modelos con BFI-44, PVQ-40, MFQ y Short Dark Triad, a temperatura 0,7, tres ejecuciones y intervalos de confianza del 95 %. Los promedios globales son 0,31 en reconstrucción semántica, 0,39 en palabras clave, F1 0,94 para asociar ítem y dimensión, MAE 0,44 al recuperar puntuaciones de opciones y MAE 0,45 al igualar objetivos; algunos modelos avanzados alcanzan MAE cercanos a 0,1–0,2. BFI y PVQ muestran resultados mayores que MFQ y SD-3. Sin embargo, las tareas revelan el texto, las dimensiones, las opciones o el objetivo, y por tanto pueden resolverse mediante inferencia semántica y razonamiento sin que el inventario estuviera en los datos de entrenamiento. La propia línea base de descripción genérica de dimensiones logra 0,32 en reconstrucción semántica, ligeramente por encima del promedio que el artículo interpreta como memorización. No hay acceso a corpus de entrenamiento, canarios, inventarios nuevos emparejados ni un diseño causal que conecte estas puntuaciones con sesgo en evaluaciones posteriores. El estudio ofrece un diagnóstico útil de familiaridad y conocimiento funcional de instrumentos, pero no identifica por sí solo contaminación ni demuestra memorización. El repositorio oficial contiene código y 19 pruebas que pasan en un entorno limpio, aunque no incluye las salidas del artículo y la reproducción integral depende de APIs y modelos mutables.

Research question

To what extent do different LLMs show familiarity with items, scoring rules, and objectives of four psychometric inventories, and can that familiarity be validly interpreted as contamination?

Method

Twenty-one models complete reconstruction, item-dimension association, score retrieval, and objective-fitting tasks on BFI-44, PVQ-40, MFQ, and SD-3. Three replicas are run at temperature 0.7, and semantic similarity, word retrieval, F1, and MAE with 95% intervals are reported. The audit additionally contrasts the tasks with their baselines and runs the tests from the official repository.

Sample: Twenty-one models, four inventaries, three runs per configuration and temperature 0.7; the article compares families and capability levels descriptively.

Findings

  • Item-dimension association is high on average (F1 0.94), while mean semantic reconstruction is 0.31.
  • Advanced models fit target scores better, with approximate MAE of 0.1-0.2 in several cases.
  • BFI-44 and PVQ-40 show greater functional familiarity than MFQ and SD-3.
  • The semantic baseline of 0.32 prevents cleanly interpreting the average of 0.31 as memorization.

Limitations

  • The tasks allow semantic inference because they expose decisive information.
  • The training corpus is not observed, nor are canaries or matched novel inventories used.
  • Three replicas are insufficient to broadly characterize stochastic variation.
  • The repository does not contain the exact outputs and depends on mutable services.

What the study does not establish

  • It does not prove that the items were in the training data.
  • It does not identify memorization as the cause over reasoning or general knowledge.
  • It does not demonstrate that the measured familiarity biases a specific psychometric evaluation.
  • It does not validate that inventory scores represent internal traits of the model.

Traceability

Scope: Full text

Version: arXiv:2510.07175v2; Findings of EACL 2026; complete 19-page PDF and TeX source; official repository commit da921a5e9b43a031e95ad005e92af92dddd2e049

Consulted source: https://arxiv.org/abs/2510.07175v2

Review: Codex full-text, visual, construct-validity and repository audit, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • 21 LLMs from major proprietary and open model families

Instruments and metrics

  • BFI-44
  • PVQ-40
  • Moral Foundations Questionnaire (MFQ)
  • Short Dark Triad (SD-3)

Data used

  • Original questionnaire items and scoring rules for four inventories

Evidence and location

  • Design, 21 models and four inventories: arXiv v2, sections 3-4 and Tables 1-2
  • Aggregate and per-inventory results: arXiv v2, section 5, Figures 2-5 and appendix tables
  • Semantic baseline and identification limits: arXiv v2, item-memorization experiments and Appendix; repository audit