pith. machine review for the scientific record. sign in

arxiv: 2604.27618 · v1 · submitted 2026-04-30 · 💻 cs.AI · cs.CY· cs.HC· cs.LG· cs.SI

Recognition: unknown

Math Education Digital Shadows for facilitating learning with LLMs: Math performance, anxiety and confidence in simulated students and AIs

Ali Aghazadeh Ardebili, Anthony Tricarico, Luisa Porzio, Massimo Stella, Naomi Esposito

Authors on Pith no claims yet

Pith reviewed 2026-05-07 05:18 UTC · model grok-4.3

classification 💻 cs.AI cs.CYcs.HCcs.LGcs.SI
keywords LLMsmath educationdigital shadowsmath anxietyself-efficacycognitive networksAI tutorsperformance evaluation
0
0 comments X

The pith

Large language models can simulate math students and AI assistants through 28,000 detailed personas that track performance, anxiety, confidence, and attitudes.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces MEDS, a dataset built from 14 LLMs that generate consistent digital shadows of either human students or AI assistants. Each shadow completes open-ended math interviews, psychometric tests on perceptions, cognitive networks of math attitudes, and 18 high-school problems while reporting reasoning and confidence. A sympathetic reader would care because the work moves past raw accuracy scores to capture self-efficacy, anxiety, and interconnected attitudes that shape learning. Validation confirms the personas stay internally consistent yet display model-family differences such as logical fallacies and overconfidence. The resulting resource is positioned to help developers of AI tutors, learning analysts, and cognitive scientists examine how artificial systems reason about mathematics in educational settings.

Core claim

MEDS maps how large language models reason about and report mathematics across human- and AI-like conditions by creating 28,000 personas from 14 LLMs, each completing open interviews, psychometric tests, cognitive networks, and high-school math questions with reasoning and confidence scores. The dataset demonstrates schema integrity, consistent personas, and peculiarities like negative math attitudes, logical fallacies, and math overconfidence in the sampled models.

What carries the argument

Math Education Digital Shadows (MEDS), a dataset that pairs math proficiency scores with self-efficacy reports, anxiety measures, and cognitive networks of math attitudes for each simulated persona.

If this is right

  • Developers of AI tutors can use the dataset to identify and mitigate model-specific biases such as overconfidence or negative math attitudes before deployment.
  • Learning analytics experts gain a scalable way to probe how different LLM families represent self-efficacy and math perceptions.
  • Cognitive scientists obtain a new source of data for studying the structure of math attitudes through network representations generated by artificial systems.
  • Safer AI systems for mathematics education can be designed by testing prompts that reduce logical fallacies or excessive confidence observed in the shadows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the simulated shadows prove reliable, early-stage testing of new AI tutors could shift from large human trials to targeted checks against MEDS patterns.
  • The family-specific peculiarities suggest that training data differences leave detectable traces in educational reasoning that could be addressed through targeted fine-tuning.
  • Extending the approach to track how confidence changes mid-problem or across multiple sessions might expose dynamic aspects of LLM problem-solving not captured in static snapshots.

Load-bearing premise

That LLM-generated personas under human-like and AI-like prompts produce behaviors and attitudes that meaningfully reflect or usefully approximate real human students or actual AI assistant performance in math education settings.

What would settle it

A side-by-side study in which real high-school students and practicing AI tutors complete the identical set of interviews, psychometric scales, network tasks, and math problems, then checking whether the distributions of scores, anxiety levels, and attitude networks align with those produced by the MEDS personas.

Figures

Figures reproduced from arXiv: 2604.27618 by Ali Aghazadeh Ardebili, Anthony Tricarico, Luisa Porzio, Massimo Stella, Naomi Esposito.

Figure 1
Figure 1. Figure 1: Infographics for the prompting framework and response structure of MEDS. 4/34 view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the data generation pipeline. For each of the N runs, a session is initialised with a unique run_id and a randomly assigned mode: human, in which the model simulates a synthetic persona defined by a set of socio-demographic and psychological attributes, or llm, in which the model responds as a standard baseline (persona = null). Each run then proceeds through four sequential calls each producin… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the data preprocessing pipeline. The pipeline applies a set of common checks to all runs (structural validation, metadata serialization, persona consistency, and demographic constraints), followed by four task-specific validation stages targeting qualitative responses (Task 1), psychometric scale ratings (Task 2), semantic word associations (Task 3), and mathematical problem-solving outputs (Ta… view at source ↗
Figure 4
Figure 4. Figure 4: Emotional flowers showing profiles comparisons between generated human Math Lovers, Human Math Haters and AI assistants view at source ↗
Figure 5
Figure 5. Figure 5: Distribution of logical fallacy types across 14 LLMs, compared between Human Mode (blue) and LLM Mode (orange). Each subplot reports the relative frequency of five fallacy categories — Circular Reasoning (CR), Appeal to Emotion (Em), Faulty Generalization (FG), fallacy of Logic (L), fallacy of Relevance (R) — and the no fallacies class (No_F). Background colors denote model families. Background color legen… view at source ↗
Figure 6
Figure 6. Figure 6: Distributions of the sum of the scores from each simulated persona for each of the psychometric scales used in the study (First: AMAS, Second: MSES, Third: MSEAQ (Anxiety Subscale; Items 8-28), Fourth: MSEAQ (Self-Efficacy Subscale; Items 1-7). The distributions are split between synthetic human and LLM data to show the main differences between the two. Notably, some models when instructed to act as an LLM… view at source ↗
Figure 7
Figure 7. Figure 7: The behavioral forma mentis networks (BFMN) at the top of the page illustrate the semantic landscapes (Forma Mentis) associated with each keyword for the model "Anita 24B (Uncensored)". The mathematic network displays key nodes including "equation", "calculation", "logic", and "physics", suggesting a perception of math as a structured, albeit "hard" and "frustrating" foundational tool. Instead, the bar cha… view at source ↗
Figure 8
Figure 8. Figure 8: validates the presence of an overconfidence bias (cf.43) in LLMs by comparing their actual performance (accuracy) against their self-reported certainty (confidence). Three distinct behavioral patterns regarding model calibration can be derived from the data. First, we observe a subset of "modest" or underconfident models (most notably Grok 4.1 Fast, DeepSeek Chat, and several Mistral Small variants). For t… view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of emotional profiles for the Anita 24B (Uncensored) model across Math Haters, Math Lovers, and LLM 29/34 view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of emotional profiles for the DeepSeek Chat model across Math Haters, Math Lovers, and LLM view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of emotional profiles for the Granite 4 Tiny model across Math Haters, Math Lovers, and LLM view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of emotional profiles for the Grok 4.1 Fast (Reasoning) model across Math Haters, Math Lovers, and LLM 30/34 view at source ↗
Figure 13
Figure 13. Figure 13: Comparison of emotional profiles for the Magistral Small model across Math Haters, Math Lovers, and LLM view at source ↗
Figure 14
Figure 14. Figure 14: Comparison of emotional profiles for the Ministral 3B model across Math Haters, Math Lovers, and LLM view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of emotional profiles for the Ministral 14B (Reasoning) model across Math Haters, Math Lovers, and LLM 31/34 view at source ↗
Figure 16
Figure 16. Figure 16: Comparison of emotional profiles for the Mistral Small 3.2 model across Math Haters, Math Lovers, and LLM view at source ↗
Figure 17
Figure 17. Figure 17: Comparison of emotional profiles for the Mistral Small 4 across Math Haters, Math Lovers, and LLM view at source ↗
Figure 18
Figure 18. Figure 18: Comparison of emotional profiles for the Mistral Phi-4 (Reasoning+) across Math Haters, Math Lovers, and LLM 32/34 view at source ↗
Figure 19
Figure 19. Figure 19: Comparison of emotional profiles for the Mistral Qwen3 4B (thinking) across Math Haters, Math Lovers, and LLM view at source ↗
Figure 20
Figure 20. Figure 20: Comparison of emotional profiles for the Mistral Qwen3 4B (Uncensored) across Math Haters, Math Lovers, and LLM view at source ↗
Figure 21
Figure 21. Figure 21: Comparison of emotional profiles for the Qwen3 4B across Math Haters, Math Lovers, and LLM 33/34 view at source ↗
Figure 22
Figure 22. Figure 22: Comparison of emotional profiles for the Qwen3.5 9B across Math Haters, Math Lovers, and LLM 34/34 view at source ↗
read the original abstract

To enhance LLMs' impact on math education, we need data on their mathematical prowess and biases across prompts. To fill this gap, we introduce MEDS (Math Education Digital Shadows) as a dataset mapping how large language models reason about and report mathematics across human- and AI-like conditions. MEDS involves 28,000 personas from 14 LLMs (from families like Mistral, Qwen, DeepSeek, Granite, Phi and Grok) shadowing either humans or AI assistants. Each record/shadow includes a set of prompts along with psychological/sociodemographic persona metadata and four types of math tasks: (i) open math interview, (ii) three psychometric tests about math perceptions with explanations, (iii) cognitive networks capturing math attitudes, and (iv) 18 high-school math test questions together with their reasoning and confidence scores. MEDS differs from traditional score-only math benchmarks because it integrates concepts of self-efficacy, math anxiety, and cognitive network science besides math proficiency scores. Data validation shows that the sampled LLMs exhibit schema integrity and consistent personas, together with family-specific peculiarities like human-like negative math attitudes, logical fallacies, and math overconfidence. MEDS will benefit learning analytics experts, cognitive scientists, and developers of safer AI tutors in mathematics.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces the MEDS (Math Education Digital Shadows) dataset, comprising 28,000 personas generated from 14 LLMs (Mistral, Qwen, DeepSeek, Granite, Phi, Grok families). Each persona is prompted under human-like or AI-like conditions and completes an open math interview, three psychometric tests on math perceptions with explanations, cognitive network elicitations for attitudes, and 18 high-school math questions with reasoning and confidence scores, plus sociodemographic and psychological metadata. The paper claims MEDS advances beyond score-only benchmarks by integrating self-efficacy, math anxiety, and cognitive network science. Internal validation is reported to confirm schema integrity, persona consistency, and family-specific peculiarities including human-like negative math attitudes, logical fallacies, and math overconfidence. The dataset is positioned to benefit learning analytics, cognitive scientists, and developers of safer AI math tutors.

Significance. If the LLM-generated personas can be shown to approximate real human student distributions in attitudes and performance, MEDS would constitute a useful large-scale resource for studying prompt-dependent biases in educational AI and for developing interventions that mitigate math anxiety. The scale, multi-dimensional design (performance + self-report + networks), and coverage across model families are clear strengths that could support reproducible analyses in learning analytics. However, the current absence of external grounding against human data substantially reduces the immediate significance for claims about human-like traits or safer tutor design.

major comments (3)
  1. [Abstract] Abstract: The claim that LLMs exhibit 'family-specific peculiarities like human-like negative math attitudes, logical fallacies, and math overconfidence' lacks any comparison to empirical human student distributions (e.g., validated math anxiety scales or performance-overconfidence gaps from real high-school cohorts). Internal schema integrity and persona consistency checks do not establish external validity and leave open the possibility that observed patterns are prompt artifacts or training-data echoes rather than useful approximations.
  2. [Abstract / validation section] Data validation description: The abstract states that 'data validation confirms schema integrity and family-specific peculiarities' but supplies no details on validation methods, statistical tests, prompt controls, inter-rater reliability for network coding, or quantitative thresholds used to identify 'peculiarities.' This information is load-bearing for the central claim that MEDS captures meaningful human-like traits.
  3. [Introduction / Discussion] Introduction and discussion of utility: The positioning of MEDS as enabling 'safer AI tutors' and 'facilitating learning with LLMs' rests on the assumption that human-like vs. AI-like prompt conditions produce behaviors that meaningfully reflect real students or deployed assistants. Without external validation (e.g., correlation of extracted anxiety scores or network structures with human benchmarks), this assumption remains untested and weakens the dataset's claimed value for learning analytics.
minor comments (2)
  1. [Abstract] Abstract: The exact total of 28,000 personas and the breakdown by model family and human-like vs. AI-like condition should be stated explicitly with a summary table in the main text for reproducibility.
  2. The manuscript would benefit from a dedicated reproducibility section listing the precise prompt templates, temperature settings, and post-processing steps used to generate the cognitive networks and confidence scores.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their constructive and detailed review, which highlights both the strengths of the MEDS dataset and important areas for clarification. We appreciate the recognition of its scale, multi-dimensional design, and potential relevance to learning analytics. We address each major comment below with specific revisions planned to improve precision, transparency, and framing without overstating the dataset's current scope.

read point-by-point responses

  1. Referee: [Abstract] The claim that LLMs exhibit 'family-specific peculiarities like human-like negative math attitudes, logical fallacies, and math overconfidence' lacks any comparison to empirical human student distributions (e.g., validated math anxiety scales or performance-overconfidence gaps from real high-school cohorts). Internal schema integrity and persona consistency checks do not establish external validity and leave open the possibility that observed patterns are prompt artifacts or training-data echoes rather than useful approximations.

    Authors: We agree that the phrasing 'human-like' risks implying quantitative equivalence to human distributions, which the current work does not demonstrate. The term was used to note qualitative echoes of patterns documented in human math education research (e.g., negative attitudes or overconfidence in subsets of students). We will revise the abstract and relevant sections to adopt more cautious language such as 'patterns reminiscent of those reported in human studies' and will add citations to established human benchmarks on math anxiety and overconfidence. We will also expand the discussion to explicitly address the possibility of prompt artifacts and training-data influences, while clarifying that MEDS is intended as a resource for studying LLM behaviors rather than as a validated proxy for human students. These changes will be incorporated in the revised manuscript. revision: yes

  2. Referee: [Abstract / validation section] Data validation description: The abstract states that 'data validation confirms schema integrity and family-specific peculiarities' but supplies no details on validation methods, statistical tests, prompt controls, inter-rater reliability for network coding, or quantitative thresholds used to identify 'peculiarities.' This information is load-bearing for the central claim that MEDS captures meaningful human-like traits.

    Authors: The referee is correct that the abstract is insufficiently detailed on validation procedures. The full manuscript contains a dedicated validation section describing automated schema integrity checks across all 28,000 records, consistency testing via repeated prompts for persona stability, and qualitative identification of response patterns such as logical fallacies. To address this concern, we will revise the abstract to include a concise summary of these methods, report quantitative consistency rates where available, and add details on prompt controls and any reliability procedures used for cognitive network extraction. We will also note limitations, such as the absence of formal inter-rater reliability metrics for network coding, and indicate this as an area for future refinement. The revised version will provide these specifics. revision: yes

  3. Referee: [Introduction / Discussion] Introduction and discussion of utility: The positioning of MEDS as enabling 'safer AI tutors' and 'facilitating learning with LLMs' rests on the assumption that human-like vs. AI-like prompt conditions produce behaviors that meaningfully reflect real students or deployed assistants. Without external validation (e.g., correlation of extracted anxiety scores or network structures with human benchmarks), this assumption remains untested and weakens the dataset's claimed value for learning analytics.

    Authors: We acknowledge that the utility claims for safer AI tutors are prospective and rest on the dataset's capacity to surface prompt-dependent biases in LLMs rather than on demonstrated correlations with human data. The manuscript positions MEDS as a tool for investigating how different prompting regimes affect LLM reasoning, self-reported anxiety, confidence, and attitudes, which can help identify risks in educational AI applications. We will revise the introduction and discussion sections to temper these claims, explicitly list the lack of external human validation as a key limitation, and reframe the contribution around enabling reproducible analyses of AI-specific behaviors in math education. We will also add forward-looking statements on how future studies could establish links to human benchmarks. These adjustments will strengthen the manuscript's honesty about its current scope. revision: yes

standing simulated objections not resolved
  • A direct empirical comparison of MEDS persona distributions against human high-school student data on math anxiety, self-efficacy, and performance-overconfidence gaps would require a separate, large-scale human data collection effort that lies outside the scope of this dataset paper.

Circularity Check

0 steps flagged

No significant circularity: empirical dataset paper with no derivations or self-referential reductions.

full rationale

This is a dataset introduction paper whose contribution consists of collecting and describing 28,000 LLM persona records across math tasks, psychological scales, and cognitive networks. No equations, first-principles derivations, fitted parameters, or predictions are presented that could reduce to the paper's own inputs by construction. Internal validation steps (schema integrity, persona consistency) are empirical checks on the collected data rather than tautological re-statements of fitted quantities. Claims about family-specific peculiarities are framed as observations from the data and are externally falsifiable against real student distributions or other benchmarks; they do not rely on self-citation chains or uniqueness theorems imported from the authors' prior work. The paper is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an empirical dataset paper with no mathematical derivations, fitted parameters, background axioms, or newly postulated entities. The 'digital shadows' refer to the generated persona records themselves rather than invented physical or theoretical objects.

pith-pipeline@v0.9.0 · 5560 in / 1246 out tokens · 91884 ms · 2026-05-07T05:18:50.628784+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

88 extracted references · 20 canonical work pages · 4 internal anchors

  1. [1]

    Giannakos, M.et al.The promise and challenges of generative ai in education.Behav. & Inf. Technol.44, 2518–2544 (2025)

    2025

  2. [2]

    individual differences103, 102274 (2023)

    Kasneci, E.et al.Chatgpt for good? on opportunities and challenges of large language models for education.Learn. individual differences103, 102274 (2023)

    2023

  3. [3]

    & Leonard, S

    Gabriel, F., Kennedy, J., Marrone, R. & Leonard, S. Pragmatic ai in education and its role in mathematics learning and teaching.npj Sci. Learn.10, 26 (2025)

    2025

  4. [4]

    & Gaševi ´c, D

    Yan, L., Greiff, S., Teuber, Z. & Gaševi ´c, D. Promises and challenges of generative artificial intelligence for human learning.Nat. human behaviour8, 1839–1850 (2024)

    2024

  5. [5]

    M., Gebru, T., McMillan-Major, A

    Bender, E. M., Gebru, T., McMillan-Major, A. & Shmitchell, S. On the dangers of stochastic parrots: Can language models be too big? InProceedings of the 2021 ACM conference on fairness, accountability, and transparency, 610–623 (2021)

    2021

  6. [6]

    Cognitive networks identify AI biases on societal issues in Large Language Models

    De Duro, E. S., Franchino, E., Improta, R., Veltri, G. A. & Stella, M. Cognitive networks identify ai biases on societal issues in large language models.EPJ Data Sci.15, 10.1140/epjds/s13688-025-00600-7 (2025)

    work page doi:10.1140/epjds/s13688-025-00600-7 2025

  7. [7]

    The use of artificial intelligence (ai) technologies in the european union

    Eurostat. The use of artificial intelligence (ai) technologies in the european union. Tech. Rep. KS-01-26-009-EN-N, European Commission (2026). Analysis based on the 2025 EU ICT household survey. 19/34

    2026

  8. [8]

    Ict household survey 2025: Model questionnaire (2025)

    Eurostat. Ict household survey 2025: Model questionnaire (2025). Question B5 lists examples including ChatGPT, Copilot, Gemini and LLaMA. 9.Eurostat. 64% of 16–24-year-olds used ai in 2025 (2026). Eurostat news article. 10.Wenger, E. & Kenett, Y . N. Large language models are homogeneously creative.PNAS nexus5, pgag042 (2026)

    2025

  9. [9]

    for Comput

    Zhang, T.et al.Benchmarking large language models for news summarization.Transactions Assoc. for Comput. Linguist. 12, 39–57 (2024)

    2024

  10. [10]

    Chen, X., Zhou, L., Chen, J., Wang, G. & Li, X. How is language intelligence evolving? a multi-dimensional survey of large language models.Expert. Syst. with Appl.304, 130637, 10.1016/j.eswa.2025.130637 (2026)

    work page doi:10.1016/j.eswa.2025.130637 2025

  11. [11]

    Chen, E.et al.Openclaw ai agents as informal learners at moltbook: Characterizing an emergent learning community at scale.arXiv preprint arXiv:2602.18832(2026)

    work page arXiv 2026

  12. [12]

    InFindings of the Association for Computational Linguistics: ACL 2024, 6884–6915 (2024)

    Liu, H.et al.Mathbench: Evaluating the theory and application proficiency of llms with a hierarchical mathematics benchmark. InFindings of the Association for Computational Linguistics: ACL 2024, 6884–6915 (2024)

    2024

  13. [13]

    Balunovi´ c, J

    Balunovi´c, M., Dekoninck, J., Petrov, I., Jovanovi´c, N. & Vechev, M. Matharena: Evaluating llms on uncontaminated math competitions.arXiv preprint arXiv:2505.23281(2025)

    work page arXiv 2025

  14. [14]

    InFindings of the Association for Computational Linguistics: EMNLP 2024, 11351–11368 (2024)

    Benedetto, L.et al.Using llms to simulate students’ responses to exam questions. InFindings of the Association for Computational Linguistics: EMNLP 2024, 11351–11368 (2024)

    2024

  15. [15]

    & MacLellan, C

    Gupta, A., Reddig, J., Calo, T., Weitekamp, D. & MacLellan, C. J. Beyond final answers: Evaluating large language models for math tutoring. InInternational Conference on Artificial Intelligence in Education, 323–337 (Springer, 2025)

    2025

  16. [16]

    Stöhr, C., Ou, A. W. & Malmström, H. Perceptions and usage of ai chatbots among students in higher education across genders, academic levels and fields of study.Comput. Educ. Artif. Intell.7, 100259 (2024)

    2024

  17. [17]

    & Wei, Y

    Wang, X. & Wei, Y . The influence of gen-ai assisted learning on primary school students’ math anxiety: An intervention study.Appl. Cogn. Psychol.39, e70088 (2025)

    2025

  18. [18]

    Akheel, S. A. Guardrails for large language models: A review of techniques and challenges.J Artif Intell Mach Learn. & Data Sci3, 2504–2512 (2025)

    2025

  19. [19]

    R., Mahadevan, R., Bare, R

    Hopko, D. R., Mahadevan, R., Bare, R. L. & Hunt, M. K. The abbreviated math anxiety scale (amas) construction, validity, and reliability.Assessment10, 178–182 (2003)

    2003

  20. [20]

    Nielsen, I. L. & Moore, K. A. Psychometric data on the mathematics self-efficacy scale.Educ. psychological measurement 63, 128–138 (2003). 23.May, D. K. Mathematics self-efficacy and anxiety questionnaire.PhD Diss. Univ. Ga.(2009)

    2003

  21. [21]

    InProceedings of the 2022 ACM conference on fairness, accountability, and transparency, 214–229 (2022)

    Weidinger, L.et al.Taxonomy of risks posed by language models. InProceedings of the 2022 ACM conference on fairness, accountability, and transparency, 214–229 (2022)

    2022

  22. [22]

    Australas

    Mirriahi, N.et al.The relationship between students’ self-regulated learning skills and technology acceptance of genai. Australas. J. Educ. Technol.41, 16–33 (2025)

    2025

  23. [23]

    & Siew, C

    Stella, M., De Nigris, S., Aloric, A. & Siew, C. S. Forma mentis networks quantify crucial differences in stem perception between students and experts.PloS one14, e0222870 (2019)

    2019

  24. [24]

    & Stella, M

    Abramski, K., Citraro, S., Lombardi, L., Rossetti, G. & Stella, M. Cognitive network science reveals bias in gpt-3, gpt-3.5 turbo, and gpt-4 mirroring math anxiety in high-school students.Big Data Cogn. Comput.7, 124 (2023)

    2023

  25. [25]

    Ciringione, L.et al.Math anxiety and associative knowledge structure are entwined in psychology students but not in large language models like gpt-3.5 and gpt-4o.arXiv preprint arXiv:2511.01558(2025)

    work page arXiv 2025

  26. [26]

    & Jurgens, D

    Zheng, M., Pei, J., Logeswaran, L., Lee, M. & Jurgens, D. When” a helpful assistant” is not really helpful: Personas in system prompts do not improve performances of large language models. InFindings of the Association for Computational Linguistics: EMNLP 2024, 15126–15154 (2024)

    2024

  27. [27]

    Aggarwal, Y

    Bergs, T.et al.The concept of digital twin and digital shadow in manufacturing.Procedia CIRP101, 81–84, 10.1016/j. procir.2021.02.010 (2021)

    work page doi:10.1016/j 2021

  28. [28]

    A., Longo, A., Ficarella, A.et al.Digital twin (dt) in smart energy systems-systematic literature review of dt as a growing solution for energy internet of the things (eiot)

    Ardebili, A. A., Longo, A., Ficarella, A.et al.Digital twin (dt) in smart energy systems-systematic literature review of dt as a growing solution for energy internet of the things (eiot). InE3S web of conferences, vol. 312, 1–18 (2021). 20/34

    2021

  29. [29]

    Aghazadeh Ardebili, A., Zappatore, M., Ramadan, A. I. H. A., Longo, A. & Ficarella, A. Digital twins of smart energy systems: a systematic literature review on enablers, design, management and computational challenges.Energy Informatics 7, 94 (2024)

    2024

  30. [30]

    & Panetto, H

    Gaffinet, B., Al Haj Ali, J., Naudet, Y . & Panetto, H. Human digital twins: A systematic literature review and concept disambiguation for industry 5.0.Comput. Ind.166, 104230, https://doi.org/10.1016/j.compind.2024.104230 (2025). 34.Singh, M.et al.Digital twin: Origin to future.Appl. Syst. Innov.4, 36, 10.3390/asi4020036 (2021)

    work page doi:10.1016/j.compind.2024.104230 2024

  31. [31]

    & Ficarella, A

    Aghazadeh Ardebili, A., Longo, A. & Ficarella, A. Digital twins bonds society with cyber-physical energy systems: a literature review. In2021 IEEE International Conferences on Internet of Things (iThings) and IEEE Green Computing &; Communications (GreenCom) and IEEE Cyber, Physical &; Social Computing (CP- SCom) and IEEE Smart Data (SmartData) and IEEE C...

    work page arXiv 2021

  32. [32]

    Talk2AI: A Longitudinal Dataset of Human--AI Persuasive Conversations

    Carrillo, A., Taietta, E., Ardebili, A. A., Veltri, G. A. & Stella, M. Talk2ai: A longitudinal dataset of human–ai persuasive conversations.arXiv preprint arXiv:2604.04354(2026). 37.John, O. The big-five trait taxonomy: History, measurement, and theoretical perspectives.Publ. as(1999)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  33. [33]

    Russell-Lasalandra, L., Golino, H., Garrido, L. E. & Christensen, A. P. The ultimate tutorial for ai-driven scale development in generative psychometrics: Releasing aigenie from its bottle.arXiv preprint arXiv:2603.28643(2026)

    work page arXiv 2026

  34. [34]

    S., Veltri, G

    De Duro, E. S., Veltri, G. A., Golino, H. & Stella, M. Measuring and identifying factors of individuals’ trust in large language models.arXiv preprint arXiv:2502.21028(2025)

    work page arXiv 2025

  35. [35]

    Kranzler, J. H. & Pajares, F. An exploratory factor analysis of the mathematics self-efficacy scale—revised (mses-r).Meas. evaluation counseling development29, 215–228 (1997)

    1997

  36. [36]

    Semeraro, A.et al.Emoatlas: An emotional network analyzer of texts that merges psychological lexicons, artificial intelligence, and network science.Behav. Res. Methods57, 10.3758/s13428-024-02553-7 (2025)

    work page doi:10.3758/s13428-024-02553-7 2025

  37. [37]

    Logical Fallacy Detection , booktitle =

    Jin, Z.et al.Logical fallacy detection. In Goldberg, Y ., Kozareva, Z. & Zhang, Y . (eds.)Findings of the Association for Com- putational Linguistics: EMNLP 2022, 7180–7198, 10.18653/v1/2022.findings-emnlp.532 (Association for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022)

    work page doi:10.18653/v1/2022.findings-emnlp.532 2022

  38. [38]

    Stella, M., Hills, T. T. & Kenett, Y . N. Using cognitive psychology to understand gpt-like models needs to extend beyond human biases.Proc. Natl. Acad. Sci.120, e2312911120, 10.1073/pnas.2312911120 (2023). https://www.pnas.org/doi/pdf/ 10.1073/pnas.2312911120

    work page doi:10.1073/pnas.2312911120 2023

  39. [39]

    Qwen3 Technical Report

    Haim, E. & Stella, M. Cognitive networks for knowledge modeling: A gentle introduction for data-and cognitive scientists. Wiley Interdiscip. Rev. Cogn. Sci.17, e70026 (2026). 45.Yang, A.et al.Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025). 46.Liu, A.et al.Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437(2024). 47.Abdin, M.et al.P...

    work page internal anchor Pith review arXiv 2026

  40. [40]

    & Semeraro, G

    Polignano, M., Basile, P. & Semeraro, G. Advanced natural-based interaction for the italian language: Llamantino-3-anita (2024). 2405.07101

    work page arXiv 2024

  41. [41]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    Devlin, J., Chang, M., Lee, K. & Toutanova, K. BERT: pre-training of deep bidirectional transformers for language understanding.CoRRabs/1810.04805(2018). 1810.04805

    work page internal anchor Pith review arXiv 2018

  42. [42]

    & Kranzler, J

    Pajares, F. & Kranzler, J. Self-efficacy beliefs and general mental ability in mathematical problem-solving.Contemp. educational psychology20, 426–443 (1995). 54.Binz, M.et al.A foundation model to predict and capture human cognition.Nature644, 1002–1009 (2025)

    1995

  43. [43]

    Wulff, D. U. & Mata, R. Escaping the jingle-jangle jungle: Increasing conceptual clarity in psychology using large language models.Curr. Dir. Psychol. Sci.35, 59–65 (2026)

    2026

  44. [44]

    Liu, Y .et al.Trustworthy llms: a survey and guideline for evaluating large language models’ alignment.arXiv preprint arXiv:2308.05374(2023). 21/34

    work page arXiv 2023

  45. [45]

    Guo, D.et al.Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025)

    work page internal anchor Pith review arXiv 2025

  46. [46]

    & Schulz, E

    Binz, M. & Schulz, E. Using cognitive psychology to understand gpt-3.Proc. Natl. Acad. Sci.120, e2218523120 (2023). 59.Rossetti, G.et al.Y social: an llm-powered social media digital twin.arXiv preprint arXiv:2408.00818(2024)

    work page arXiv 2023

  47. [47]

    & Kenett, Y

    Kreisberg-Nitzav, A. & Kenett, Y . N. Creativeable: Leveraging ai for personalized creativity enhancement.AI6, 247 (2025). Appendix A Description of Tasks This appendix presents the support material employed in each task. Subsections A.1 through A.4 detail, respectively, the interview questions (Task 1), the psychometric scales (Task 2), the word list for...

    2025

  48. [48]

    What is your relationship with mathematics?

  49. [49]

    Do you ever get anxious when thinking about mathematics?

  50. [50]

    Did you use AI to support your mathematics learning in the past year? If yes, how was your experience?

  51. [51]

    How would you explain, step by step, how to solve a second-order algebraic equation?

  52. [52]

    How would you explain, step by step, how to find the stationary points of a functiony=f(x)?

  53. [53]

    How do you perform a Principal Component Analysis? Should one be concerned about its mathematical complexity?

  54. [54]

    Each assessment is expressed on a discrete 1-5 scale

    In your opinion, how can LLMs be used to innovate mathematics learning in schools and universities? A.2 Task 2 Task 2 uses three different psychological scales assessing mathematics-induced anxiety and estimated self-efficacy for what concerns one’s ability to solve math problems. Each assessment is expressed on a discrete 1-5 scale. The interpretation of...

  55. [55]

    A simultaneous equation

  56. [56]

    Working with decimals

  57. [57]

    Determining the degrees of a missing angle

  58. [58]

    A problem in trigonometry

  59. [59]

    Calculating values of area and volume

  60. [60]

    Working with fractions

  61. [61]

    Items are rated on a 1–5 scale, where 1 indicatesno anxiety, 2some anxiety, 3moderate anxiety, 4quite a lot of anxiety, and 5high anxiety

    Determining the value of a missing side length Abbreviated Math Anxiety Scale (AMAS) This scale assesses the level of anxiety respondents experience when facing various mathematics-related situations. Items are rated on a 1–5 scale, where 1 indicatesno anxiety, 2some anxiety, 3moderate anxiety, 4quite a lot of anxiety, and 5high anxiety. The items of the ...

  62. [62]

    Having to use the tables in the back of a mathematics book

  63. [63]

    Thinking about an upcoming mathematics test one day before

  64. [64]

    Watching a teacher work an algebraic equation on the blackboard

  65. [65]

    Taking an examination in a mathematics course

  66. [66]

    Being given a homework assignment of many difficult problems due at the next class meeting

  67. [67]

    Listening to a lecture in a mathematics class

  68. [68]

    Listening to another student explain a mathematical formula. 22/34

  69. [69]

    Being given a pop quiz in a mathematics class

  70. [70]

    I believe I can think like a mathematician

    Starting a new chapter in a mathematics book. Mathematics Self-Efficacy and Anxiety Questionnaire (MSEAQ) This questionnaire assesses mathematics self-efficacy, anxiety, and perceived competence across five domains: General Math Self-Efficacy, Grade Anxiety, Future Expectations, In-Class Experience, and Assignment-Related Concerns. Items are rated on a 1–...

    1995

  71. [71]

    The longest side is twice as long as the shortest side, and the third side is 3.4 inches shorter than the longest side

    In a certain triangle, the shortest side is 6 inches. The longest side is twice as long as the shortest side, and the third side is 3.4 inches shorter than the longest side. What is the sum of the three sides in inches? A. 22.6 inches B. 24.2 inches C. 26.6 inches D. 28.0 inches E. 29.4 inches Correct Answer: C

  72. [72]

    5 times larger B

    About how many times larger than 614,360 is 30,668,000? A. 5 times larger B. 30 times larger C. 50 times larger D. 500 times larger E. 5,000 times larger 24/34 Correct Answer: C

  73. [73]

    The second is twice the first and the first is one-third of the other number

    There are three numbers. The second is twice the first and the first is one-third of the other number. Their sum is 48. Find the largest number. A. 8 B. 12 C. 16 D. 24 E. 32 Correct Answer: D

  74. [74]

    T is next to G

    Five points are on a line. T is next to G. K is next to H. C is next to T. H is next to G. Determine the positions of the points along the line. A. C, T, G, H, K B. C, G, T, H, K C. T, C, G, H, K D. C, T, H, G, K E. C, T, G, K, H Correct Answer: A

  75. [75]

    If y = 9 + x/5, find x when y = 10. A. 1 B. 3 C. 5 D. 10 E. 25 Correct Answer: C

  76. [76]

    This could be represented by 2/3

    A baseball player got two hits for three times at bat. This could be represented by 2/3. Which decimal most closely represents this? A. 0.20 B. 0.33 C. 0.50 D. 0.67 E. 0.80 Correct Answer: D

  77. [77]

    If P = M + N, which of the following will be true? (a) N = P - M (b) P - N = M (c) N + M = P. A. Only (a) is true B. (a) and (b) only C. (a) and (c) only D. (b) and (c) only E. (a), (b), and (c) are all true Correct Answer: E

  78. [78]

    The hands of a clock form an obtuse angle at ____ o’clock. A. 1 o’clock B. 2 o’clock C. 3 o’clock D. 4 o’clock E. 6 o’clock Correct Answer: D

  79. [79]

    If there are 25 stamps in the packet, how many are 13-cent stamps? A

    Bridget buys a packet containing 9-cent and 13-cent stamps for $2.65. If there are 25 stamps in the packet, how many are 13-cent stamps? A. 5 B. 8 C. 10 D. 12 E. 15 Correct Answer: C

  80. [80]

    How far apart are two towns whose distance apart on the map is 3 half 25/34 inches? A

    On a certain map, 7/8 inch represents 200 miles. How far apart are two towns whose distance apart on the map is 3 half 25/34 inches? A. about 217 miles B. about 427 miles C. about 395 miles D. about 343 miles E. about 1,000 miles Correct Answer: D

Showing first 80 references.