Systematic Review Flow (PRISMA 2020)
Transparent flow of article identification, LLM screening, full-text eligibility assessment, and inclusion into living synthesis with expert validation.
✓ PRISMA funnel conservation verified (100% stage closure)
Zero data leakage: n_in = n_out + ∑n_excl + n_pending verified across identification, screening, eligibility, and synthesis.
ID: 77 →
Screen: 73 →
Eligible: 76 →
Included: 7
Expert Validation & Taste Loop
Independent of the formal PRISMA selection rules: editorial verdicts by the curator to calibrate quality and active learning.
👍 0
Approved in digest
(0 total in corpus)
👎 1
Rejected with reason
(0 pre / 0 post analysis)
Breakdown of expert rejection reasons:
duplicate: 1
📋 Structured Review Summary (PRISMA 2020: Item 2)
Background: Living systematic review of Large Language Models (LLMs) and autonomous agents in higher education. Focuses on engineering countermeasures to academic dishonesty and redesign of assessment.
Objectives: Systematic identification, screening, and synthesis of empirical evidence regarding reliability, methodological rigor, and consensus challenges in higher education GenAI integration.
Eligibility Criteria: Scholarly works with open-access full text; topic relevance R ≥ 4.0; weighted final score ≥ 5.0.
Information Sources: Multi-register tracking: arXiv, Crossref, OpenAlex, Semantic Scholar, ERIC, plus forward/backward citation loops (depth 1–2).
Risk of Bias & Quality: Multi-criteria methodological rigor (M, 0–10). Screening reliability validated via dual evaluation: Gwet's AC1, stratified recall, and ROC AUC.
Synthesis & Included: Included n = 7 verified works into the living review with extracted claims, limitations, and empirical findings.
🎯 PICO / PECO Search & Eligibility Framework (PRISMA 2020: Items 4–5)
| Component | Protocol Definition | Inclusion / Exclusion Boundaries |
|---|---|---|
| P (Population) | Higher education: universities, colleges, undergraduate/graduate students, instructors, and examiners. | Included: higher/tertiary education. Excluded: purely K-12 primary education without transferability. |
| I / E (Intervention / Exposure) | Use of modern LLMs (GPT-4, Claude, Gemini, Llama) and agentic frameworks for exam solving, tutoring, or automated grading. | Included: empirical testing and deployment of LLMs. Excluded: legacy non-generative ML or symbolic AI. |
| C (Comparator) | Traditional non-AI assessments, human baselines, or comparative evaluation of engineering defenses (guardrails, multi-agent validation). | Included: baselines, cohort comparisons, or countermeasures. Excluded: purely speculative essays without comparator. |
| O (Outcomes) | Task performance, exam breach rates, countermeasure resilience, shift in assessment rubrics, cognitive learning outcomes. | Included: empirical accuracy, resilience, rubric difficulty, psychometrics. Excluded: unverified claims. |
| S (Study Design) | Empirical studies: controlled experiments (RCT), quasi-experiments, case studies, psychometric evaluations, benchmark audits. | Included: peer-reviewed/preprint research with empirical data. Excluded: opinion blogs, press releases. |
* Footnote *: Prefilter Mathematical Model, Indicators & Scoring Formula (Rubric v2.3)
Rubric Indicators & Parameters
| Symbol | Indicator Name | Weight / Value |
|---|---|---|
| M | Methodological Rigor | wM = 0.32 |
| E | Engineering Countermeasures | wE = 0.25 |
| C | Consensus Challenge | wC = 0.43 |
| N | Novelty | wN = 0.00 |
| R | Topic Relevance (0–10) | Rmin = 4.0, Rfull = 7.0 |
| Gate | Binary Scope Gate | PASS / FAIL |
| Поріг | Pass Threshold | Score ≥ 5.0 |
Relevance R is a multiplier, not a summand: papers with R < R_min receive 0, and high-quality off-topic papers cannot outrank topical ones.
Scoring Formulae (MathML Core)
1. Final Prefilter Score:
2. Weighted Quality Score (Q):
3. Logical Prefilter Screening Criterion:
Methods & reporting (PRISMA 2020 items 1–15, 23–27)
Review type, objective, eligibility (1, 3–5)
Living systematic review: a daily-updated, automation-assisted selection. It is a literature-selection tool, not a published review with meta-analysis.
Topic (given to the model): Validity of educational assessment under generative AI and how to keep it (AI-resistant / AI-resilient assessment). Core: how generative AI changes what grades and test scores mean (validity arguments in the sense of Kane and Messick), assessment redesign, and quantitative methods that detect or model the distortion in assessment and learning data: item response theory, differential item functioning between humans and LLMs, item parameter drift, rater drift, item compromise and preknowledge, score shifts under unsupervised or online conditions (proctored vs unproctored exams, cheating, grade inflation), change-point and concept-drift detection, LLMs as test takers or raters, computerized adaptive testing and automated test assembly. Computational, statistical and engineering work is preferred, but a solid conceptual or empirical education study of assessment under generative AI belongs here too. Detectors of AI-generated text and watermarking are only adjacent to the topic. The audience is university educators and researchers: work about primary and secondary school (K-12) is off topic unless its findings or methods clearly transfer to higher education.
- Include: a scholarly article with topic relevance R ≥ 4.0, final score ≥ 5.0; a legal open-access full text that was actually read.
- Exclude: non-articles (datasets, code, errata), duplicates and preprints of published works, retracted works, K-12 with no transfer to higher education, works without full text (paywall, fetch errors), expert-rejected works (👎, reason recorded).
- Manual additions (origin = manual) bypass score screening but go through eligibility assessment.
Information sources & last searched (6)
| Channel | Records | Last record |
|---|---|---|
search — arXiv, OpenAlex, Semantic Scholar, ERIC | 11616 | 2026-10-29 10:22 |
ref — reference lists of read works | 1649 | 2026-10-29 10:25 |
manual — added manually | 27 | 2026-10-07 12:05 |
Citation candidates take metadata from OpenAlex; Crossref/Unpaywall supply DOIs, versions and open PDFs.
Search strategy (7)
Current search version: v1.0. A work matches if its title/abstract hits a term from EVERY group of at least one query (OR within a group; case-insensitive, word boundaries).
Queries (groups):
education × ai; psychometrics × ai; integrity × education; drift × education; drift × psychometrics
education: education, educational, student, learner, teacher, lecturer, classroom, university, higher education, school, curriculum, assessment, exam, examination, examinee, test taker, test score, grading, homework, coursework, online course, essay, academic integrity, academic misconduct, plagiarism, cheating, e-learning, MOOC
ai: generative AI, GenAI, large language model, LLM, ChatGPT, GPT-4, chatbot, AI-generated, AI-assisted, AI-resistant, AI-resilient, AI tool, artificial intelligence
integrity: proctoring, proctored, unproctored, non-proctored, remote proctoring, online exam, online test, online testing, exam integrity, testing integrity, academic integrity, academic dishonesty, cheating, contract cheating, grade inflation
psychometrics: item response theory, IRT, Rasch, psychometric, educational measurement, educational testing, differential item functioning, computerized adaptive testing, adaptive testing, automated test assembly, knowledge tracing, automated scoring, automated essay scoring, validity argument, construct validity, assessment validity, predictive validity
drift: parameter drift, dynamic item response, item parameter drift, rater drift, severity drift, concept drift, change point, change-point detection, item compromise, preknowledge, item exposure, test security, aberrant response, person-fit, differential item functioning
Semantic Scholar phrases: assessment validity generative AI; AI-resilient assessment higher education; AI-resistant assessment design; generative AI exam performance natural experiment; differential item functioning large language models humans; LLM as test taker item response theory; proctored versus unproctored online exam scores; generative AI grade inflation; item parameter drift detection; item compromise preknowledge detection; concept drift knowledge tracing; change point detection educational testing
arXiv: cs.CY, cs.CL, cs.HC, cs.AI, cs.LG, stat.AP, stat.ME. Result caps per query: OpenAlex 50, Semantic Scholar 25, ERIC 25. Citation searching: depth 1 from read and 👍 works, depth 2 from 👍 only; neighbours of 👍 works bypass the lexical filter.
| Search version | Records (search) | First | Last |
|---|---|---|---|
v1.0 | 11616 | 2026-10-06 | 2026-10-29 |
Selection process & automation (8, 9, 11)
- Abstract screening: a single automated reviewer — an LLM (gemini-3.1-flash-lite, gemini-3.5-flash-lite), batches of 20, models alternate; every evaluation is stored (prefilter_evaluations) with model and rubric version.
- Full-text analysis and data collection: gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash, prompt
v2.1. Only after PDF→Markdown conversion; the model is never called without full text. - Verification: code checks every quote against the text (exact match after normalisation); unconfirmed claims are discarded.
- Human: no independent second reviewer. The owner gives 👍/👎 verdicts on a sample (50% near threshold, 30% top, 20% random); they steer the queue and few-shot examples but do not change the score.
- Data items (10): summary, method/data/sample, claims, limitations, study design (assessment_design, psychometric_modeling, natural_experiment, detector_evaluation, survey, rct, quasi_experiment, qualitative, theoretical, review, other), tags, rubric criteria.
- Risk of bias (11): not assessed formally (RoB, GRADE not applied); criterion M (methodological rigor, 0–10) is a ranking score, not a RoB tool. No effect measures or meta-analysis (12, 13d–f, 15): synthesis is qualitative, via article analyses and tags.
| Screening model | Scored |
|---|---|
gemini-3.1-flash-lite | 6322 |
gemini-3.5-flash-lite | 5786 |
— | 459 |
| Analysis model | prompt | Analyses |
|---|---|---|
gemini-3.1-flash-lite | v2.1 | 275 |
gemini-3.6-flash | v2.1 | 102 |
gemini-3.8-flash | v2.1 | 86 |
gemini-3.7-flash | v2.1 | 80 |
Scoring rubric v2.3 (0–10 anchors)
The model scores 0–10 per criterion; the anchors below are embedded in the prompt. Formula and weights are in the footnote above. Changing anchors or weights without a new version is blocked by a hash check.
| Criterion | Weight | 2 | 5 | 8 |
|---|---|---|---|---|
| Topic relevance (R) | × min(1, R/7.0) | No link to education or assessment: image, audio or model watermarking, model IP protection, general NLP benchmarks, AI in industry. | Adjacent: generative AI in education without an assessment focus; academic integrity with no evidence on assessment outcomes; detectors of AI-generated text or of student AI use; general psychometrics (adaptive testing, test assembly, IRT) with no drift, validity or AI angle. | Core: validity of assessment under generative AI (what grades and scores mean when AI is available); how grades shift under unsupervised conditions or cheating (proctored vs unproctored exams, grade inflation); AI-resistant or AI-resilient assessment design; quantitative methods that detect or model distorted assessment or learning data (DIF between humans and LLMs, item parameter or rater drift, item compromise/preknowledge, change-point or concept-drift detection, LLMs as test takers or raters). |
| Methodological rigor (M) | 0.32 | Vague evaluation, convenience sample of a few dozen, self-report only, no control group or baseline. | Clear dataset or sample (hundreds), standard metrics or validated instruments, comparison with baselines or a control group. | Randomized/controlled or preregistered, large or multi-site/multilingual data, adversarial or robustness checks, effect sizes with uncertainty. |
| Engineering countermeasures (E) | 0.25 | No method or countermeasure: describes impact, attitudes or policy only. | A concrete method or countermeasure with an evaluation: statistical test or model, detector, scoring procedure, or an assessment design tried on real cohorts. | A rigorous, reusable method: formal model with guarantees or error control, open code, validated on both real and simulated data. |
| Consensus challenge (C) | 0.43 | Reconfirms standard wisdom. | Nuances the consensus (e.g. detectors fail under mild paraphrasing; AI use does not harm an outcome it is assumed to harm). | Directly refutes high-profile claims with rigorous evidence. |
| Novelty (N) | 0.00 | Replicates well-known findings (another survey of student attitudes to ChatGPT, another test of one detector on a handful of prompts). | New dataset, benchmark, population or setting, or a previously untried technique. | Fundamentally new attack or defence, a theoretical/impossibility result, or a finding that changes how the field frames the problem. |
The model interpolates between anchors. The deep-analysis score is computed by code with the same formula from criteria_breakdown.
| Rubric version | Papers scored |
|---|---|
v2.0 | 261 |
v2.1 | 183 |
v2.2 | 20 |
v2.3 (current) | 12103 |
Screening Validation: Reliability & Agreement Metrics
Reference standard: expert verdicts. Full PRISMA systematic review metrics computed: recall, specificity, NPV, F₁, F₂, ROC AUC, DOR, paradox-immune Gwet's AC₁, and continuous score agreement.
| Measure | Value | Meaning |
|---|---|---|
| Seed-set recall | 50% (1/2), 95% CI 9–91% | Share of manually added 👍 works (bypassing screening) scoring ≥ threshold 5.0. Unbiased sensitivity estimate on this set. |
| Estimated corpus recall | 100% (100–100%) | Post-stratified by score band from judged works; conservative interval. Relevant works missed ≈ 0 (0–0). Not measured (too few verdicts in band): 10549 screened-out works. |
| Sensitivity / Recall | 100% (12/12), 95% CI 76–100% | Share of relevant works passed by screening: TP / (TP + FN). |
| Specificity | 20% (4/20), 95% CI 8–42% | Share of irrelevant works rejected by screening: TN / (TN + FP). |
| Precision / PPV | 43% (12/28), 95% CI 27–61% | Share of expert-approved among screening passes: TP / (TP + FP). |
| Negative Predictive Value (NPV) | 100% (4/4), 95% CI 51–100% | Share of truly irrelevant among screened-out works: TN / (TN + FN). Protects against losing relevant papers. |
| F-scores (F₁ and F₂) | F₁: 0.60 · F₂: 0.79 | F₂ weights recall twice as much as precision — critical safety metric against false exclusions in PRISMA. |
| Balanced Accuracy | 60.0% | Arithmetic mean of sensitivity and specificity, unaffected by class imbalance. |
| Diagnostic Odds Ratio (DOR) | 6.82 | DOR = (TP · TN) / (FP · FN) with Haldane-Anscombe correction. Overall discriminative capability of screening. |
| ROC AUC | 0.65 (95% CI 0.43–0.84) | Probability a 👍 work outscores a 👎 one (0.5 = chance). CI computed via DeLong / Bootstrap. |
| Inter-Rater Agreement (Gwet's AC₁ & Cohen's κ) | — | Gwet's AC₁ is paradox-immune inter-rater agreement on double-scored works, robust to severe class imbalance (80-90% screening rejection). |
| Continuous Score Agreement |
MAD: — r Пірсона: — |
Mean Absolute Difference (MAD) and Pearson r between scores of screening models on shared papers. |
| Score band | Works | Judged | 👍 | 👍 rate |
|---|---|---|---|---|
| 0.0–2.0 | 6729 | 2 (too few) | 0 | 0% |
| 2.0–4.0 | 2910 | 0 (too few) | 0 | — |
| 4.0–5.0 | 910 | 2 (too few) | 0 | 0% |
| 5.0–6.0 ✓ | 1072 | 3 | 2 | 67% |
| 6.0–8.0 ✓ | 856 | 23 | 8 | 35% |
| 8.0–10.0 ✓ | 27 | 2 (too few) | 2 | 100% |
| Screening model | Seed recall | AUC | Sens / Spec | Prec / NPV | F₁ / AC₁ | 👍 / 👎 (TP/FP/FN/TN) |
|---|---|---|---|---|---|---|
gemini-3.1-flash-lite |
50% (1/2), 95% CI 9–91% | 0.63 | 100% (12/12), 95% CI 76–100% / 16% (3/19), 95% CI 6–38% | 43% (12/28), 95% CI 27–61% / 100% (3/3), 95% CI 44–100% | 0.60 / 0.05 | 12 / 19 (12/16/0/3) |
gemini-3.5-flash-lite |
— | — | — / 100% (1/1), 95% CI 21–100% | — / 100% (1/1), 95% CI 21–100% | — / 1.00 | 0 / 1 (0/0/0/1) |
— |
— | — | — / — | — / — | — / — | 0 / 0 (0/0/0/0) |
To narrow intervals, judge works from lower score bands: /admin/taste includes 20% random picks. Metrics update automatically with each verdict.
Methodological Rigor & Study Designs (11, 18)
Methodological rigor M (0–10) is evaluated during full-text analysis. Study design distribution reflects verified empirical classifications.
Mean Rigor (M)
7.21 / 10
Analyzed papers: 76
Quality Stratification
🟢 High rigor (M ≥ 7.0): 56
🟡 Moderate rigor (4.0 ≤ M < 7.0): 19
🔴 Low rigor (M < 4.0): 1
| Study Design | Paper Count |
|---|---|
psychometric_modeling | 65 |
assessment_design | 7 |
quasi_experiment | 2 |
theoretical | 1 |
review | 1 |
PRISMA 2020 27-Item Compliance Checklist
Full audit of compliance with PRISMA 2020 checklist (Page et al., 2021) for automated systematic reviews.
| № | Section / Item | Checklist Description | Status | Implementation |
|---|---|---|---|---|
| 1 | Title | Identify as living systematic review | ✓ Впроваджено | Page title, report metadata |
| 2 | Abstract | Structured summary (background, objectives, methods, results) | ✓ Впроваджено | #prisma-abstract block above |
| 3 | Rationale | Rationale in context of existing knowledge | ✓ Впроваджено | Topic description and rationale |
| 4 | Objectives | Explicit objectives using PICO / PECO | ✓ Впроваджено | #pico-framework table above |
| 5 | Eligibility criteria | Inclusion/exclusion criteria, language, dates | ✓ Впроваджено | Thresholds R, Score, Open Access |
| 6 | Information sources | Databases, registers, search dates | ✓ Впроваджено | arXiv, Crossref, OpenAlex, Semantic Scholar, ERIC, citations |
| 7 | Search strategy | Full search strategy, boolean filters, versions | ✓ Впроваджено | search_version, term groups, API limits |
| 8 | Selection process | Selection process, automation, independent screening | ✓ Впроваджено | LLM screening rotation + editorial taste loop |
| 9 | Data collection | Data collection from full-text reports | ✓ Впроваджено | PDF → MD parser, deep prompt v2.1 |
| 10a-b | Data items | List of outcomes and study characteristics | ✓ Впроваджено | Summary, sample, design, claims, countermeasures |
| 11 | Study risk of bias | Assessment of study methodological quality | ✓ Впроваджено | Rigor criterion M (0–10), quality distribution |
| 12 | Effect measures | Effect measures for outcomes | — Обґрунтовано N/A | Qualitative synthesis without meta-analytic pooled effects |
| 13a-f | Synthesis methods | Synthesis criteria and grouping methods | ✓ Впроваджено | Tag-based clustering, comparative study cards |
| 14 | Reporting bias | Assessment of reporting/publication bias | ✓ Впроваджено | Analysis of paywall exclusions & open access skew |
| 15 | Certainty assessment | Certainty assessment of body of evidence | ✓ Впроваджено | Automated quote grounding against full-text Markdown |
| 16 | Study selection | Selection results (flow diagram, counts) | ✓ Впроваджено | 100% stage conservation closure verified |
| 17 | Study characteristics | Characteristics of included studies table | ✓ Впроваджено | Interactive table with DOI, venue, year, tags |
| 18 | Risk of bias in studies | Risk of bias findings across studies | ✓ Впроваджено | Methodological rigor M & study designs table |
| 19 | Results of individual studies | Results for each individual study | ✓ Впроваджено | Public paper pages /doi/{doi} with full analysis |
| 20a-d | Results of syntheses | Results of syntheses & sub-groups | ✓ Впроваджено | Daily living editions, RSS feed, tag filters |
| 21 | Reporting biases | Results of reporting biases assessment | ✓ Впроваджено | Paywall n=0, retracted n=0 |
| 22 | Certainty of evidence | Summary of evidence certainty | ✓ Впроваджено | Gwet's AC₁, Cohen's κ, stratified recall, ROC AUC |
| 23a-d | Discussion | Discussion, limitations, implications | ✓ Впроваджено | Limitations section (23b, 23c) in methods |
| 24a-c | Registration & protocol | Registration, protocol availability, amendments | ✓ Впроваджено | Public repo, version history of search & rubric |
| 25 | Support | Financial & non-financial support | ✓ Впроваджено | Free API tiers declared in methods |
| 26 | Competing interests | Competing interests statement | ✓ Впроваджено | Declared in methods section |
| 27 | Data & code availability | Availability of data, analytical code, materials | ✓ Впроваджено | Export to CSV / JSON buttons, open-source code |
Limitations (23b, 23c)
- Open access only: paywalled papers never reach the synthesis, so the sample skews to open and preprint work.
- Screening ranks by quality and novelty rather than pure eligibility, so the sample skews towards counter-consensus work. Validity is measured against one expert (see above); a single LLM reviewer, model bias and version differences remain possible.
- The lexical filter and result caps can miss works with other terminology; sources skew English-language.
- Records scored under an older rubric version are not rescored automatically (see version table).
- Free Gemini quota: daily limits slow analysis; some selected works wait in the queue.
Registration, protocol, amendments, support, data (24–27)
- Registration: not registered. The protocol is the configuration (topic, queries, rubric, thresholds) in the project repository; it is reproduced in this block.
- Amendments: search versions (search_version) and rubric versions (rubric_version), tables above.
- Funding: none declared, free API tiers are used. Competing interests: none declared.
- Data: analyses and metadata are on article pages, RSS and sitemap; full texts are private (copyright) and not published.
Filter by Tags
Select one or more tags to refine the flow numbers and article corpus.
ℹ️ <strong>Note on thematic filtering:</strong> Thematic tags are generated during full-text analysis or assigned manually by the editor. Tag filtering focuses the table and flow on that marked subset.
#psychometrics (95)
#Psychometrics (52)
#Item Response Theory (38)
×
#academic integrity (38)
#item response theory (33)
#assessment validity (28)
#Assessment Validity (25)
#IRT (24)
#higher education (24)
#construct validity (23)
#Automated Essay Scoring (20)
#Generative AI (19)
#automated essay scoring (19)
#Construct Validity (18)
#LLM as rater (16)
#assessment_design (16)
#generative AI (16)
#Academic Integrity (14)
#Automated Scoring (14)
#Higher Education (14)
#LLM Evaluation (13)
#Large Language Models (13)
#assessment design (13)
#automated assessment (13)
#AI-resilient assessment (11)
#academic_integrity (11)
#assessment redesign (11)
#assessment-validity (11)
#assessment_validity (11)
#benchmarking (11)
#large language models (11)
#automated scoring (10)
#item_response_theory (10)
#validity (10)
#ChatGPT (9)
#DIF (9)
#GPT-4 (9)
#Inter-Rater Reliability (9)
#LLM as Rater (9)
#LLM evaluation (9)
#LLM-as-a-judge (9)
#LLMs as Test Takers (9)
#concept drift (9)
#differential item functioning (9)
#AI-resistant assessment (8)
#Assessment Design (8)
#Knowledge Tracing (8)
#LLMs as raters (8)
#Prompt Engineering (8)
#construct_validity (8)
#prompt engineering (8)
#rater drift (8)
#AI-Resilient Assessment (7)
#Assessment Redesign (7)
#Automated Item Generation (7)
#Computer Science Education (7)
#Computerized Adaptive Testing (7)
#Validity (7)
#computerized adaptive testing (7)
#higher_education (7)
Reset tags
| # | Status / Stage | Paper | LLM Score | Expert Review |
|---|---|---|---|---|
| 1 | In Buffer |
The Phantom Hierarchy: How Scientifically Invalid AI IQ Scores Acquire Authority and Cause Real Harm
Zenodo (CERN European Organization for Nuclear Research) (2026) · doi:10.5281/zenodo.18849483 · v1.0 |
7.0 (gemini-3.6-flash) | — |
| 2 | In Buffer |
K-12EduBench: A Benchmark for Evaluating Large Language Models’ Knowledge, Problem-Solving, and Educational Goal Cognition in K-12 Education
Proceedings of the AAAI Conference on Artificial Intelligence (2026) · doi:10.1609/aaai.v40i40.40744 · v1.0 |
3.8 (gemini-3.6-flash) | — |
| 3 | In Buffer |
From Diagnosis to Generalization: A Cognitive Approach to Data Selection for Educational LLMs
Proceedings of the AAAI Conference on Artificial Intelligence (2026) · doi:10.1609/aaai.v40i26.39300 · v1.0 |
7.0 (gemini-3.6-flash) | — |
| 4 | In Buffer |
Assessing the Quality of AI-Generated Exams: A Large-Scale Field Study
Proceedings of the AAAI Conference on Artificial Intelligence (2026) · doi:10.1609/aaai.v40i45.41205 · v1.0 |
6.6 (gemini-3.8-flash) | — |
| 5 | In Buffer |
Automated Estimation of Difficulty Levels in Math Word Problems using Linguistic, Mathematical, and Semantic Features
International Journal of Mathematical Engineering and Management Sciences (2026) · doi:10.33889/ijmems.2026.11.2.028 · v1.0 |
3.2 (gemini-3.6-flash) | — |
| 6 | Included |
Algorithmic drift in generative adaptive learning
arXiv (2026) · doi:10.48550/arxiv.2610.31041 · v1.0 |
7.3 (gemini-3.7-flash) | — |
| 7 | In Buffer |
Evaluating Large Language Models through Multidimensional Item Response Theory: A Comprehensive Case Study on ENEM
Brazilian Symposium in Information and Human Language Technology (2025) · doi:10.5753/stil.2025.37846 · v1.0 |
7.2 (gemini-3.7-flash) | — |
| 8 | In Buffer |
Assisting the grading of a handwritten general chemistry exam with artificial intelligence
Discover Artificial Intelligence (2026) · doi:10.1007/s44163-026-01606-4 · v1.0 |
6.1 (gemini-3.7-flash) | — |
| 9 | In Buffer |
GEN-EXAM: AN AUTONOMOUS ADAPTIVE ASSESSMENT SYSTEM BASED ON LLM-DRIVEN DYNAMIC QUESTION GENERATION AND COMPLEXITY ANALYSIS
International Journal of Scientia (2026) · doi:10.66194/gvncijsv1i1-006 · v1.0 |
1.9 (gemini-3.8-flash) | — |
| 10 | In Buffer |
AI-Based Question Paper Generation and Quality Optimization: An Integrated Computational Approach to Intelligent Assessment Construction
Indian Journal of Computer Science and Technology (2026) · doi:10.59256/indjcst.20260502085 · v1.0 |
2.6 (gemini-3.7-flash) | — |
| 11 | In Buffer |
A Measurement Perspective on Benchmarking LLM-Generated Text Detectors using Item Response Theory
arXiv (2026) · doi:10.48550/arxiv.2610.24972 · v1.0 |
4.0 (gemini-3.8-flash) | — |
| 12 | In Buffer |
Lost in Benchmarks? Rethinking Large Language Model Benchmarking with Item Response Theory
Proceedings of the AAAI Conference on Artificial Intelligence (2026) · doi:10.1609/aaai.v40i41.40814 · v1.0 |
6.7 (gemini-3.7-flash) | — |
| 13 | In Buffer |
Easy2Hard-Bench: Standardized Difficulty Labels for Profiling LLM Performance and Generalization
neural information processing systems (2024) · doi:10.52202/079017-1407 · v1.0 |
3.9 (gemini-3.8-flash) | — |
| 14 | In Buffer |
Does a Language Model Have a Personality? What Four Studies Say, and What Measurement Should Do Next
Zenodo (CERN European Organization for Nuclear Research) (2026) · doi:10.5281/zenodo.21865128 · v1.0 |
4.4 (gemini-3.7-flash) | — |
| 15 | In Buffer |
Benchmark Collapse
Zenodo (CERN European Organization for Nuclear Research) (2026) · doi:10.5281/zenodo.21939754 · v1.0 |
5.1 (gemini-3.8-flash) | — |
| 16 | In Buffer |
The Catastrophic Paradox of Human Cognitive Frameworks in Large Language Model Evaluation: A Comprehensive Empirical Analysis of the CHC-LLM Incompatibility
arXiv (2025) · doi:10.48550/arxiv.2511.18302 · v1.0 |
2.5 (gemini-3.7-flash) | — |
| 17 | In Buffer |
Latency-Response Theory Model: Evaluating Large Language Models via Response Accuracy and Chain-of-Thought Length
arXiv (2025) · doi:10.48550/arxiv.2512.07019 · v1.0 |
6.5 (gemini-3.7-flash) | — |
| 18 | In Buffer |
Reconstructing Item Characteristic Curves using Fine-Tuned Large Language Models
arXiv (2026) · doi:10.48550/arxiv.2601.02580 · v1.0 |
5.1 (gemini-3.7-flash) | — |
| 19 | In Buffer |
Synthetic Student Responses: LLM-Extracted Features for IRT Difficulty Parameter Estimation
arXiv (2026) · doi:10.48550/arxiv.2602.00034 · v1.0 |
6.1 (gemini-3.6-flash) | — |
| 20 | In Buffer |
Take Out Your Calculators: Estimating the Real Difficulty of Question Items with LLM Student Simulations
arXiv (2026) · doi:10.48550/arxiv.2601.09953 · v1.0 |
6.9 (gemini-3.8-flash) | — |
| 21 | Included |
Diagnosing the Reliability of LLM-as-a-Judge via Item Response Theory
arXiv (2026) · doi:10.48550/arxiv.2602.00521 · v1.0 |
7.7 (gemini-3.7-flash) | — |
| 22 | In Buffer |
FoundationalASSIST: An Educational Dataset for Foundational Knowledge Tracing and Pedagogical Grounding of LLMs
arXiv (2026) · doi:10.48550/arxiv.2602.00070 · v1.0 |
4.9 (gemini-3.8-flash) | — |
| 23 | In Buffer |
Confident Rankings with Fewer Items: Adaptive LLM Evaluation with Continuous Scores
arXiv (2026) · doi:10.48550/arxiv.2601.13885 · v1.0 |
5.2 (gemini-3.7-flash) | — |
| 24 | In Buffer |
Estimating Exam Item Difficulty with LLMs: A Benchmark on Brazil's ENEM Corpus
arXiv (2026) · doi:10.48550/arxiv.2602.06631 · v1.0 |
6.7 (gemini-3.7-flash) | — |
| 25 | In Buffer |
Quantifying and Mitigating Socially Desirable Responding in LLMs: A Desirability-Matched Graded Forced-Choice Psychometric Study
arXiv (2026) · doi:10.18653/v1/2026.acl-long.1865 · v1.0 |
5.5 (gemini-3.7-flash) | — |
| 26 | In Buffer |
Quantifying construct validity in large language model evaluations
arXiv (2026) · doi:10.48550/arxiv.2602.15532 · v1.0 |
6.3 (gemini-3.6-flash) | — |
| 27 | In Buffer |
RankLLM: Weighted Ranking of LLMs by Quantifying Question Difficulty
arXiv (2026) · doi:10.48550/arxiv.2602.12424 · v1.0 |
4.3 (gemini-3.7-flash) | — |
| 28 | In Buffer |
Evaluating Cross-Modal Reasoning Ability and Problem Characteristics with Multimodal Item Response Theory
arXiv (2026) · doi:10.48550/arxiv.2603.02663 · v1.0 |
6.2 (gemini-3.6-flash) | — |
| 29 | In Buffer |
Leveraging Computerized Adaptive Testing for Cost-effective Evaluation of Large Language Models in Medical Benchmarking
arXiv (2026) · doi:10.48550/arxiv.2603.23506 · v1.0 |
7.0 (gemini-3.7-flash) | — |
| 30 | In Buffer |
Ranking Reasoning LLMs under Test-Time Scaling
arXiv (2026) · doi:10.48550/arxiv.2603.10960 · v1.0 |
3.7 (gemini-3.6-flash) | — |
| 31 | In Buffer |
Agent psychometrics: Task-level performance prediction in agentic coding benchmarks
arXiv (2026) · doi:10.48550/arxiv.2604.00594 · v1.0 |
5.3 (gemini-3.8-flash) | — |
| 32 | In Buffer |
Which AI model is truly the best: how reliable are public benchmarks of frontier AI models?
Zenodo (CERN European Organization for Nuclear Research) (2026) · doi:10.5281/zenodo.23007588 · v1.0 |
0.0 (gemini-3.8-flash) | — |
| 33 | In Buffer |
Rubric-Calibrated Preferences: Cross-Query Calibration of LLM Judgments via Item Response Theory
arXiv (Cornell University) (2026) · doi:10.48550/arxiv.2609.35739 · v1.0 |
0.0 (gemini-3.7-flash) | — |
| 34 | In Buffer |
Toward Interpretable Automated Item Generation: Evaluating Q-Matrix Representations of Item Difficulty Using LLTM
Practical assessment, research & evaluation (2026) · doi:10.7275/pare.3653 · v1.0 |
3.7 (gemini-3.7-flash) | — |
| 35 | In Buffer |
Growing Pains: Extensible and Efficient LLM Benchmarking Via Fixed Parameter Calibration
arXiv (2026) · doi:10.48550/arxiv.2604.12843 · v1.0 |
0.0 (gemini-3.7-flash) | — |
| 36 | In Buffer |
MCQ Difficulty Prediction via Modeling Learner Heterogeneity Using Data-Driven Cognitive Profiling
arXiv (2026) · doi:10.48550/arxiv.2605.16290 · v1.0 |
6.2 (gemini-3.7-flash) | — |
| 37 | In Buffer |
Stress-Testing the Reasoning Competence of LLMs With Proofs Under Minimal Formalism
arXiv (2026) · doi:10.48550/arxiv.2605.12524 · v1.0 |
5.5 (gemini-3.6-flash) | — |
| 38 | In Buffer |
Interpretable Difficulty-Aware Knowledge Tracing in Tutor-Student Dialogues
arXiv (2026) · doi:10.48550/arxiv.2605.01097 · v1.0 |
3.6 (gemini-3.6-flash) | — |
| 39 | In Buffer |
Estimating LLM Grading Ability and Response Difficulty in Automatic Short Answer Grading via Item Response Theory
arXiv (2026) · doi:10.48550/arxiv.2605.00238 · v1.0 |
6.9 (gemini-3.7-flash) | — |
| 40 | In Buffer |
Which AI model is truly the best: how reliable are public benchmarks of frontier AI models?
Zenodo (CERN European Organization for Nuclear Research) (2026) · doi:10.5281/zenodo.23036392 · v1.0 |
5.1 (gemini-3.8-flash) | — |
| 41 | Included |
An Interpretable and Scalable Framework for Evaluating Large Language Models
arXiv (2026) · doi:10.48550/arxiv.2605.07046 · v1.0 |
7.5 (gemini-3.7-flash) | — |
| 42 | In Buffer |
Automated Scoring of Scientific Creativity in German
Journal of Creative Behavior (2024) · doi:10.1002/jocb.658 · v1.0 |
3.4 (gemini-3.8-flash) | — |
| 43 | Included |
Auditing LLM Benchmarks with Item Response Theory
arXiv (2026) · doi:10.48550/arxiv.2605.30504 · v1.0 |
7.3 (gemini-3.6-flash) | — |
| 44 | In Buffer |
Cognitive Episodes in LLM Reasoning Traces Enable Interpretable Human Item Difficulty Prediction
arXiv (2026) · doi:10.48550/arxiv.2606.28186 · v1.0 |
6.8 (gemini-3.7-flash) | — |
| 45 | In Buffer |
The Easy Trap: Why LLMs Underestimate Misconception-Driven Difficulty
arXiv (2026) · doi:10.48550/arxiv.2607.26067 · v1.0 |
5.7 (gemini-3.8-flash) | — |
| 46 | In Buffer |
Latent Confidence Alignment for LLM Self-Assessment
arXiv (2026) · doi:10.48550/arxiv.2606.21937 · v1.0 |
6.5 (gemini-3.6-flash) | — |
| 47 | In Buffer |
AgentCAT: Simulating Computerized Adaptive Testing via Multi-Agent Large Language Models
arXiv (2026) · doi:10.48550/arxiv.2606.21832 · v1.0 |
3.7 (gemini-3.7-flash) | — |
| 48 | In Buffer |
LLMs Struggle to Measure What Distinguishes Students of Different Proficiency Levels: A Study of Item Discrimination in Reading Comprehension Assessment
arXiv (2026) · doi:10.48550/arxiv.2606.18709 · v1.0 |
6.4 (gemini-3.6-flash) | — |
| 49 | Included |
Extending Item Response Theory for Efficient and Meaningful Multilingual Evaluation
arXiv (2026) · doi:10.48550/arxiv.2606.15643 · v1.0 |
7.2 (gemini-3.7-flash) | — |
| 50 | In Buffer |
Assessment of School Performance Through a Multilevel Latent Markov Rasch Model
Journal of Educational and Behavioral Statistics (2011) · doi:10.3102/1076998610381396 · v1.0 |
0.0 (gemini-3.8-flash) | — |