AI-accelerated End-to-End Framework for Rapid Professional Upskilling

# AI-accelerated End-to-End Framework for Rapid Professional Upskilling
Source: [https://arxiv.org/html/2607.14044](https://arxiv.org/html/2607.14044)
###### Abstract

By 2030, 59 of every 100 workers will need reskilling or upskilling, yet the average time to close an enterprise skills gap grew from roughly 3 days in 2014 to 36 days in 2018\. Most current frameworks accelerate single stages of upskilling programs and generally lack industry validation\. We present an end\-to\-end framework that applies AI acceleration across five stages of knowledge acquisition, content development, content review and verification, teaching, and assessment development; with a strong focus on both production and learning efficiency\. Three strong external signals validates the framework: the US National Association of State Boards of Accountancy reviewed and approved an upskilling program built on the framework for continuing\-professional\-education credits; 3 learners followed the program and passed the NVIDIA Certified Professional in Agentic AI exam in a significantly short amount of time, with 14 more in progress; the program’s knowledge base supports complex downstream analysis such as the production of a robust 1,267 risk item dataset for managing multi\-agent AI system risks\.

## IIntroduction

Fifty\-nine of every 100 workers will need reskilling or upskilling by 2030, and 11 of them are unlikely to receive it\. In other words, roughly 120 million workers are at the medium\-term risk of redundancy\[[37](https://arxiv.org/html/2607.14044#bib.bib1)\]\. The same forecast expects 39% of core skills to change or become outdated by 2030 as 170 million jobs are created and 92 million displaced\[[37](https://arxiv.org/html/2607.14044#bib.bib1)\]\. The skills themselves decay\. Skill relevance averages about five years, while that of technical skills is about two and a half\[[11](https://arxiv.org/html/2607.14044#bib.bib2)\]\. Organizational response is moving the other way\. The average time to close an enterprise skills gap through classroom or online training rose from roughly 3 days in 2014 to 36 days in 2018\[[18](https://arxiv.org/html/2607.14044#bib.bib3)\]\. The result is a mismatch paradox as layoffs coexist with unfilled vacancies because required competencies change faster than workers adapt\. The binding constraint is no longer whether to upskill but how fast rigorous upskilling programs can be produced and executed\.

Generative AI sharpens the problem because its benefits are uneven\. An assistant deployed to 5,179 customer\-support agents raised issues resolved per hour by 15% on average but by 30% for less skilled and less experienced workers\[[5](https://arxiv.org/html/2607.14044#bib.bib4)\]; among 453 professionals, ChatGPT helped initially weaker writers most\[[24](https://arxiv.org/html/2607.14044#bib.bib5)\]; and a field experiment with 758 consultants found gains concentrated among lower\-scoring participants\[[12](https://arxiv.org/html/2607.14044#bib.bib6)\]\. We interpret this pattern through a two\-regime framing of the cited results\. When work reduces to fixed prompts with easy\-to\-judge outputs, AI acts as a*leveler*and novices match experts\. When work demands flexible prompting and carries a high evaluation burden, AI acts as a*multiplier*and expertise compounds\. Therefore, how a workforce is upskilled shapes which regime it occupies\.

The stakes are institutional as much as individual\. In a survey of 116 executives at large organizations, roughly one in four lacked a clear view of how automation will affect skill requirements, and nearly one\-third doubted their HR infrastructure could execute a skills strategy\[[14](https://arxiv.org/html/2607.14044#bib.bib7)\]\. Institutions that cannot produce rigorous, judgment\-building upskilling programs at the pace of their AI exposure will absorb that exposure unprepared\.

We present a framework that closes this production\-capability gap by applying AI acceleration to every stage of upskilling\-program rather than to just one or few single stage\. As Section II details, existing frameworks accelerate few single stages and generally lack external industry validation\. We are explicit about the nature of our evidence: we report design, artifacts, and external validation signals, not controlled\-comparison evidence\. The paper contributes \(C1\) an end\-to\-end AI\-accelerated production pipeline for certification\-grade programs \(aligned to a reputable vendor’s certification blueprint and sufficient to prepare candidates for its exam\) spanning knowledge acquisition, content development, content review and verification, AI\-tutor coaching, and assessment development; \(C2\) a dual\-efficiency design pairing AI\-accelerated production with learning\-efficient outputs; \(C3\) specific external validation signals including: approval by NASBA \(the National Association of State Boards of Accountancy\) for continuing\-professional\-education \(CPE\) credits; certification passes \(3 learners, 14 in progress\); downstream capability validated by institutional subject\-matter\-expert \(SME\) reviews; and \(C4\) documented stage\-level methods enabling replication\. Section II reviews related work and gaps, Section III describes the framework’s five stages, Section IV reports the validation signals, Section V discusses implications, and Section VI concludes\.

## IIBackground and Related Work

Adult professionals do not learn the way children do\. Knowles characterizes adults as self\-directed, experience\-anchored, problem\-centered learners who require instruction built around authentic professional tasks\[[17](https://arxiv.org/html/2607.14044#bib.bib12)\]\. Yet AI educational technologies are designed and evaluated predominantly for K\-12 contexts and remain “poorly aligned” with adult learners’ needs\[[26](https://arxiv.org/html/2607.14044#bib.bib13)\], and the LLM planning tools targeting self\-directed learners face documented transparency and hallucination problems\[[8](https://arxiv.org/html/2607.14044#bib.bib14)\]\.

Our framework is built different\. It begins with our fundamental understanding and our definition of*rapid upskilling*\- minimizing time\-to\-competency for professionals acquiring a frontier technical topic \- where competency is demonstrated against an external standard rather than self\-reported\.

### II\-AFrameworks for Rapid Upskilling

Existing frameworks accelerate one of four loci: content delivery, content production, the learning process, or the program\. Hence, their evidence is correspondingly partial\.

Delivery\-side approaches compress instruction into short units, but the gains fade\. In a controlled comparison, ChatGPT and Google users outperformed controls on immediate lower\-order tasks, yet retention scores converged to an e\-textbook group’s\[[1](https://arxiv.org/html/2607.14044#bib.bib15)\]\. Compressed delivery therefore accelerates exposure, not competency\.

Production\-side frameworks automate the creation of materials\. Instructional Agents generates end\-to\-end course materials with a multi\-agent LLM pipeline\[[39](https://arxiv.org/html/2607.14044#bib.bib16)\]; a related system casts LLM agents as learning designers in a five\-stage pipeline\[[34](https://arxiv.org/html/2607.14044#bib.bib17)\]; and ARCHED, a human\-centered pipeline guided by Bloom’s taxonomy, produced learning objectives statistically indistinguishable from expert\-written ones\[[19](https://arxiv.org/html/2607.14044#bib.bib18)\]\. However, none verifies content at the claim level, faces learners, or validates against an external certification standard\.

Learning\-side frameworks accelerate the learner rather than the materials\. RAG\-PRISM delivers adaptive retrieval\-augmented tutoring for rapid workforce development\[[25](https://arxiv.org/html/2607.14044#bib.bib19)\]\. Tutor CoPilot, a randomized controlled trial of a human\-AI system in live tutoring, raised topic mastery by 4 percentage points, and by 9 for students of lower\-rated tutors\[[35](https://arxiv.org/html/2607.14044#bib.bib22)\]\. Deployed systems corroborate the approach: a classroom dialogue system kept 71% of conversations fully on\-track\[[20](https://arxiv.org/html/2607.14044#bib.bib23)\], and LLM\-generated feedback in introductory programming raised the odds of eventual correctness\[[13](https://arxiv.org/html/2607.14044#bib.bib24)\]\. Learner\-side acceleration works but presupposes existing materials\.

Program\-level efforts span the lifecycle with human labor: the AI Technicians program \- a rapid occupational AI training with the US Army \- succeeded under conditions its authors call “a cost that other organizations may not be able to bear”\[[28](https://arxiv.org/html/2607.14044#bib.bib20)\]\. A systematic review of LLMs in programming education confirms the field remains a collection of point solutions\[[41](https://arxiv.org/html/2607.14044#bib.bib21)\]\.

### II\-BGap Analysis

Our survey exposes four gaps\.

First, the landscape is fragmented\. For example, production pipelines stop at materials\[[39](https://arxiv.org/html/2607.14044#bib.bib16),[34](https://arxiv.org/html/2607.14044#bib.bib17),[19](https://arxiv.org/html/2607.14044#bib.bib18)\]\. No surveyed framework covers end\-to\-end results, from knowledge acquisition through successful industry assessment\. Section III presents an integrated pipeline covering all five stages\.

Second, verification is missing where it is most needed\. Averaged across models, LLMs hallucinated on 60\.33% of post\-knowledge\-cutoff questions\[[2](https://arxiv.org/html/2607.14044#bib.bib25)\]\. Detection is feasible: FAVA pairs a hallucination taxonomy with a retrieval\-augmented detector\[[21](https://arxiv.org/html/2607.14044#bib.bib26)\], and VeriScore verifies extracted claims with a fine\-tuned verifier\[[30](https://arxiv.org/html/2607.14044#bib.bib27)\]\. Yet no surveyed education pipeline incorporates such a layer\. Section III\-C answers with a dedicated verification stage\.

Third, default LLM pedagogy is shallow\. In a 223\-domain tutoring testbed, LLMs produced correct next\-step actions only 52–70% of the time\[[36](https://arxiv.org/html/2607.14044#bib.bib31)\]; the deployed successes above engineered pedagogy deliberately\[[35](https://arxiv.org/html/2607.14044#bib.bib22),[13](https://arxiv.org/html/2607.14044#bib.bib24)\]\. Section III\-D answers with explicitly designed coaching protocols\.

Fourth, outcomes are rarely measured against anything external, let alone industry certification exams on frontier topics\. A few reasons include: GenAI evaluation practice lacks scientific rigor\[[33](https://arxiv.org/html/2607.14044#bib.bib28)\], zero\-shot prompting underperformed traditional NLP in curricular analytics\[[38](https://arxiv.org/html/2607.14044#bib.bib29)\], and circular LLM\-as\-judge validation is a documented validity threat\[[31](https://arxiv.org/html/2607.14044#bib.bib32)\]\. A framework that generates its own success measure proves nothing\. Sections III\-E and IV answer with blueprint\-aligned assessment and validation against an external certification body\.

## IIIThe Crew Scaler Framework

The framework organizes rapid upskilling as a five\-stage pipeline: knowledge acquisition, content development, content review and verification, AI\-tutor teaching, and assessment development \(Fig\.[1](https://arxiv.org/html/2607.14044#S3.F1)\)\. The pipeline is rapid for two reinforcing reasons\.*Production efficiency*means AI compresses the time required to produce instructional materials at every stage\.*Learning efficiency*means the outputs themselves are structured to reduce learner time\-to\-competency through prerequisite ordering, spaced review, misconception\-keyed distractors, and adaptive tutoring\. Humans retain the roles where judgment carries the most weight such as blueprint design, subject\-matter\-expert \(SME\) review, misconception authoring, and item rating\. In the mean time, AI absorbs the volume work those judgments govern, keeping human expertise in the multiplier regime described in Section I\. Each stage therefore pairs an AI\-acceleration mechanism with a learning\-efficiency mechanism and a quality\-control check \(Table[I](https://arxiv.org/html/2607.14044#S3.T1)\)\. Finally, the pipeline’s outputs face external validation the authors do not control \(Fig\.[1](https://arxiv.org/html/2607.14044#S3.F1), bottom\)\.

KnowledgeAcquisitionContentDevelopmentContentReview &VerificationAI\-TutorCoachingAssessmentDevelopmentHuman judgment: blueprints, SME review, misconceptions, item ratingsExternal checks: NCP\-AAI exam; risk\-taxonomy derivation; federal SME review; NASBA CPE reviewFigure 1:The five\-stage pipeline\. AI accelerates production within each stage; humans retain high\-judgment roles \(top\); outputs face external checks \(bottom\); all learner\-facing stages draw on the stage\-one knowledge substrate\. NCP\-AAI: NVIDIA Certified Professional in Agentic AI\.TABLE I:Stage\-level summary: each stage pairs an AI\-acceleration mechanism with a learning\-efficiency mechanism and a quality control### III\-AKnowledge Acquisition

We surveyed 27 literature\-survey for general knowledge domain and 26 threat\-modeling methods targeting a specialized knowledge domain before designing this stage\. None was simultaneously scalable and evidence\-grounded\[[23](https://arxiv.org/html/2607.14044#bib.bib42)\]\. That gap motivated our*Knowledge Domain Exploring Guide*\(KDEG\), a four\-step procedure: domain framing, practice\-grounded job\-task analysis, blueprinting into weighted content domains with a table of specifications, and evidence mapping that links each objective to authoritative references\. AI accelerates the two labor\-dominated steps of domain exploration and knowledge\-item extraction against the blueprint, while humans retain framing, weighting, and evidence admission\. Quality control is structural rather than post hoc\. Specifically, extracted content occupies a four\-level dependency hierarchy \- foundational concepts, primary building blocks, integrated concepts, and applied or advanced material \- linked by strict dependency chains\. Because material arrives in prerequisite order, the hierarchy bounds the context a learner needs at each step to concepts already mastered, and coverage gaps surface against the blueprint rather than through inspection\. Applied to an emerging and uniquely complex knowledge domain of multi\-agent\-systems, the stage produced the approximately 3,000\-page knowledge base\.

### III\-BContent Development

Knowledge base chapters originate as AI\-generated drafts and pass through three disciplined transformations\. Structural drafting aligns each draft to the human currated knowledge domain blueprint through a fixed skeleton \(Overview, Learning Objectives, Content, Key Concepts, Assessment\) carrying 3 to 8 SMART learning objectives per chapter, hands\-on labs, and exam\-style practice items\. A condensation pass applies named compression techniques such as minimalist documentation, information mapping, worked\-example fading rather than ad hoc cutting\. AI performs the bulk drafting while the fixed templates hold the pedagogical shape constant, so human effort concentrates on judging fit against the blueprint\. A beginner\-readability pass then enforces rules diagnosed from defects in actual drafts\. For example, the “One New Element” rule permits exactly one new complexity per section, and cumulative review distributes roughly 70% current\-chapter, 20% prior\-chapter, and 10% foundational questions, operationalizing spaced retrieval practice\[[27](https://arxiv.org/html/2607.14044#bib.bib39)\]\. A six\-pass revision checklist closes each chapter\.

### III\-CContent Review and Verification

LLMs hallucinate on a majority of questions past their knowledge cutoff\[[2](https://arxiv.org/html/2607.14044#bib.bib25)\]\. AI\-drafted content therefore requires its own verification machinery\. The framework specifies three layers: automated detection, expert human review, and immutable audit documentation\. Layer 1 adapts the RAGAS framework into a four\-part accuracy standard of faithfulness, context precision, answer relevancy, and harmfulness\. To further improve issue detection, the approach uses a four\-type hallucination taxonomy: factual, reasoning, contextual, and true fabrications\. This supports research showing that both fine\-grained taxonomies and claim\-level ’extract\-then\-verify’ pipelines enhance accuracy\[[21](https://arxiv.org/html/2607.14044#bib.bib26),[30](https://arxiv.org/html/2607.14044#bib.bib27)\]\. We executed Layer 1 as a scripted pass producing six reports scored against explicit numeric targets, and the results were mixed: platform accuracy scored 95% on a 20\-example sample and cross\-chapter integrity found 0 invalid references among 268 checked; conversely, only 5 of 10 chapters met the 70% pedagogical\-progression bar, and assessment materials stood at 63% of the assessment\-bank target at that time\. We report this mixed picture deliberately: criterion\-referenced checks that can fail \(and did\) are evidence that the instrument measures rather than ratifies\[[33](https://arxiv.org/html/2607.14044#bib.bib28)\]\. Both gaps entered a tracked remediation backlog; the assessment gap was subsequently closed \(the completed 530\-question bank postdates that pass\), while the pedagogical\-progression remediation remains tracked\. Layer 2 SME sign\-off is fully specified but not yet fully evidenced in execution records\.

### III\-DAI\-Tutor Coaching

Step\-based tutoring approaches human\-tutor effect sizes \(d=0\.76d=0\.76\)\[[32](https://arxiv.org/html/2607.14044#bib.bib38)\], yet in a 223\-domain testbed current LLMs labeled incorrect learner actions no better than chance\[[36](https://arxiv.org/html/2607.14044#bib.bib31)\], and recent work narrows such gaps by training tutors against student\-outcome objectives\[[29](https://arxiv.org/html/2607.14044#bib.bib30)\]\. We therefore specified the tutor’s pedagogy in advance as a library of 16 named protocols rather than a single general\-purpose prompt, grounded in the retrieval\-practice, productive\-failure, and affect literatures\[[27](https://arxiv.org/html/2607.14044#bib.bib39),[16](https://arxiv.org/html/2607.14044#bib.bib37),[10](https://arxiv.org/html/2607.14044#bib.bib41)\]\. These protocols include: direct explanation, Socratic questioning, worked examples, hint escalation, spaced retrieval, productive failure, affective support that prioritizes boredom over frustration, and an academic\-integrity guardrail\. A selection layer picks the active protocol each turn by fixed priority: integrity risk first, then affective signals, then group context, then learner intent\. Six cross\-cutting principles constrain every protocol: knowledge grounding, learner modeling \(formalizable through knowledge tracing\[[9](https://arxiv.org/html/2607.14044#bib.bib40)\]\), verify\-before\-trusting, ask\-don’t\-tell, calibrated intervention\[[4](https://arxiv.org/html/2607.14044#bib.bib33)\], and step\-level feedback\. Error diagnosis serves learning efficiency\. Specifically, the tutor classifies a wrong answer as conceptual, procedural, factual, or careless, then checks it against a per\-chapter misconception catalog, so feedback names the specific misconception rather than a generic error\.

### III\-EAssessment Development

In a field study across 91 classes, AI\-generated questions performed comparably to expert\-created standardized\-exam items under item\-response\-theory analysis\[[15](https://arxiv.org/html/2607.14044#bib.bib34)\]\. Our pipeline pursues that standard by construction\. Every item traces to an atomic, ID\-tagged knowledge item that records at least two documented misconceptions at extraction time, and distractors are engineered from those misconceptions rather than invented when the question is written\. Stem plans map difficulty to coverage\. Specifically, easy stems test one knowledge item, medium stems combine two to three, and hard stems synthesize three to five\. We fix the difficulty distribution up front at 30% easy, 50% medium, 20% hard, consistent with evidence that explicit difficulty control steers LLM\-generated item pools away from trivially easy questions\[[40](https://arxiv.org/html/2607.14044#bib.bib36)\]\. Misconception\-keyed distractors target exactly the property that simulation\-based item analysis measures as distractor efficiency\[[22](https://arxiv.org/html/2607.14044#bib.bib35)\]\. The pipeline produced a 530\-question assessment bank tagged to a 10\-domain, 53\-skill blueprint, three 100\-question mock exam forms comprising 300 unique items, 96 chapter quizzes, and ten 70\-question simulated practice tests\. Quizzes supply low\-stakes retrieval practice while blueprint\-matched tests rehearse certification conditions, and both trace to the same misconception library\.

## IVResults

We report three validation signals, ordered from the most direct evidence: certification outcomes, capability outcomes, and independent accreditation\. The signals are independent of one another, each is externally checkable, and none is self\-graded which is a property prior AI\-accelerated instructional pipelines have generally lacked\[[39](https://arxiv.org/html/2607.14044#bib.bib16)\]\. Because the deployment is early while involves emerging topics, we state each sample size plainly and draw no claim beyond what the observed numbers support\.

### IV\-ACertification Outcomes

Three learners prepared for the NVIDIA Certified Professional in Agentic AI \(NCP\-AAI\) exam using only the framework’s knowledge base\. All three passed, a 100% pass rate to date withn=3n=3\. The exam is administered and scored by Nvidia’s appointed certification vendor\. We control neither its content nor its grading\. We report these outcomes as observed and do not extrapolate the pass rate to larger cohorts\. Three passes establish that studying the knowledge base alone can carry a learner through a vendor\-scored professional certification\. They do not establish the rate at which it will do so\. Fourteen additional learners are progressing toward the exam within the complete program, and we will report their outcomes as they accrue\. Notably, the Nvidia Agentic AI certification \(NCP\-AAI\) is a new certification with currently limited education resources\. Most learners are not even aware that this certification exists\.

### IV\-BCapability Outcomes

The knowledge base also served as the direct input to a systematic risk analysis of a baseline multi\-agent AI system\[[23](https://arxiv.org/html/2607.14044#bib.bib42)\]\. In that analysis, threat\-modeling agents worked through the roughly 3,000\-page knowledge base chapter by chapter, and produced 1,267 risk items across 81 categories and 14 domains\. This signal tests a property that certification exams do not: whether the knowledge base is complete and well\-structured enough to support downstream expert\-level analysis, rather than exam preparation alone\. The multi\-agent system risk item dataset passed surface validation when it was presented in front of around 500 US federal employees and is being peered reviewed in a historically awarded A\* journal\.

### IV\-CThird\-Party Accreditation

Finally, NASBA reviewed an education program built on the framework’s outputs and approved it for CPE credits\. For context, NASBA National Registry review assesses program design and delivery against established CPE standards\. It is independent of the authors, the certification vendor, and the adopting agencies\. This signal tests the program as an educational product before a standards body with no stake in the framework, its vendor alignment, or its adopters\.

Taken together, the three signals triangulate the framework from unaffiliated external directions; each is limited on its own, and their evidential value lies in their convergence across independent external parties\.

## VDiscussion

These signals matter because of timing\. Technical skills carry a half\-life of roughly two and a half years\[[11](https://arxiv.org/html/2607.14044#bib.bib2)\], and educational materials for frontier topics are structurally scarce\. When a certification is new such as the NCP\-AAI, preparation corpora, assessment banks, and trained instructor pools are scarce or absent\. Conventional instructional\-design cycles build these artifacts sequentially\. For frontier content the cycle runs slower than the content’s decay, and taught programs add a further stage because instructors must themselves be trained before they can teach\. An AI\-accelerated end\-to\-end pipeline changes this constraint\. Inside the window in which the knowledge is still current, our framework produced a roughly 3,000\-page prerequisite\-ordered knowledge base, verified chapters, 16 tutoring protocols, and a 530\-question assessment bank tagged to a 10\-domain, 53\-skill blueprint\. This is a significant improvement considering prior systems accelerate single stages of this chain\. Notably, any unaccelerated stage re\-imposes the original bottleneck on the entire chain\.

We also offer an interpretation of why acceleration at this scale did not cost the framework its grounding\. Humans keep the high\-judgment roles in blueprint design, SME review, misconception authoring, and item rating, while AI absorbs the volume work of drafting, condensation, and cross\-referencing\. This division mirrors the leveling pattern in the studies of Section I\.

The stakes extend beyond individual learners\. Beyond the executive readiness gap noted in Section I\[[14](https://arxiv.org/html/2607.14044#bib.bib7)\], national\-scale exposure far exceeds visible adoption\[[7](https://arxiv.org/html/2607.14044#bib.bib8)\]\. Certification capacity in frontier topics is one mechanism by which that latent exposure converts into workforce readiness rather than displacement\. The conversion matters for distribution as well as output for reasons such as: automation’s inequality effects are often non\-monotonic and can rise as technology approaches top human skill\[[3](https://arxiv.org/html/2607.14044#bib.bib9)\]; economic value increasingly bifurcates toward verification\-grade expertise; and entry\-level pressure is already visible in a reported 16% relative employment decline for workers aged 22 to 25 in AI\-exposed occupations\[[6](https://arxiv.org/html/2607.14044#bib.bib10)\]\.

We therefore read the framework not as a convenience for exam candidates but as infrastructure \- a way for institutions and economies to build upskilling capacity in new technical domains at the speed those domains now change\.

## VIConclusion

We presented an end\-to\-end, AI\-accelerated, and learning\-efficient pipeline that produced a certification\-grade training program for a frontier topic, multi\-agent AI systems, supported by strong external signals\. All three learners who studied only the knowledge base passed the NVIDIA Certified Professional in Agentic AI exam\. The same knowledge base seeded a threat\-modeling analysis yielding 1,267 risk items\[[23](https://arxiv.org/html/2607.14044#bib.bib42)\]; fand NASBA approved an upskilling program built on these outputs for CPE credits\. For organizations facing training gaps in topics too new to have textbooks, instructors, or item banks, the framework offers a documented path to produce all three while the knowledge is current\.

## References

- \[1\]M\. Akgun and S\. Toker\(2025\)Short\-term gains, long\-term gaps: the impact of GenAI and search technologies on retention\.InProc\. AIED 2025 Posters and Late Breaking Results,Communications in Computer and Information Science, Vol\.2591,pp\. 44–52\.External Links:[Document](https://dx.doi.org/10.1007/978-3-031-99264-3%5F6)Cited by:[§II\-A](https://arxiv.org/html/2607.14044#S2.SS1.p2.1)\.
- \[2\]A\. Alessa, P\. Somane, A\. T\. Lakshminarasimhan, J\. Skirzynski, J\. McAuley, and J\. M\. Echterhoff\(2025\-12\)Quantifying cognitive bias induction in LLM\-generated content\.InProc\. IJCNLP\-AACL,Mumbai, India,pp\. 2890–2910\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.ijcnlp-long.155)Cited by:[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p3.1),[§III\-C](https://arxiv.org/html/2607.14044#S3.SS3.p1.1)\.
- \[3\]S\. G\. Benzell and K\. R\. Myers\(2026\-01\)Automation experiments and inequality\.Working PaperTechnical Report34668,National Bureau of Economic Research\.Note:Also available as arXiv:2510\.24923External Links:[Document](https://dx.doi.org/10.3386/w34668)Cited by:[§V](https://arxiv.org/html/2607.14044#S5.p3.1)\.
- \[4\]C\. Borchers, A\. Gurung, Q\. Liu, D\. R\. Thomas, M\. Khalil, and K\. R\. Koedinger\(2026\)Brief but impactful: how human tutoring interactions shape engagement in online learning\.InProc\. 16th Int\. Learning Analytics and Knowledge Conf\. \(LAK\),Bergen, Norway\.External Links:[Document](https://dx.doi.org/10.1145/3785022.3785041)Cited by:[§III\-D](https://arxiv.org/html/2607.14044#S3.SS4.p1.1)\.
- \[5\]E\. Brynjolfsson, D\. Li, and L\. R\. Raymond\(2025\)Generative AI at work\.The Quarterly Journal of Economics140\(2\),pp\. 889–942\.External Links:[Document](https://dx.doi.org/10.1093/qje/qjae044)Cited by:[§I](https://arxiv.org/html/2607.14044#S1.p2.1)\.
- \[6\]C\. Catalini, X\. Hui, and J\. Wu\(2026\)Some simple economics of AGI\.External Links:2602\.20946,[Link](https://arxiv.org/abs/2602.20946)Cited by:[§V](https://arxiv.org/html/2607.14044#S5.p3.1)\.
- \[7\]A\. Chopraet al\.\(2025\)The iceberg index: measuring skills\-centered exposure in the AI economy\.External Links:2510\.25137,[Document](https://dx.doi.org/10.48550/arXiv.2510.25137)Cited by:[§V](https://arxiv.org/html/2607.14044#S5.p3.1)\.
- \[8\]J\. Chun, Y\. Zhao, H\. Chen, and M\. Xia\(2025\)PlanGlow: personalized study planning with an explainable and controllable LLM\-driven system\.InProc\. 12th ACM Conf\. Learning @ Scale \(L@S\),Palermo, Italy,pp\. 116–127\.External Links:[Document](https://dx.doi.org/10.1145/3698205.3729541)Cited by:[§II](https://arxiv.org/html/2607.14044#S2.p1.1)\.
- \[9\]A\. T\. Corbett and J\. R\. Anderson\(1994\)Knowledge tracing: modeling the acquisition of procedural knowledge\.User Modeling and User\-Adapted Interaction4\(4\),pp\. 253–278\.External Links:[Document](https://dx.doi.org/10.1007/BF01099821)Cited by:[§III\-D](https://arxiv.org/html/2607.14044#S3.SS4.p1.1)\.
- \[10\]S\. D’Mello and A\. Graesser\(2012\)Dynamics of affective states during complex learning\.Learning and Instruction22\(2\),pp\. 145–157\.External Links:[Document](https://dx.doi.org/10.1016/j.learninstruc.2011.10.001)Cited by:[§III\-D](https://arxiv.org/html/2607.14044#S3.SS4.p1.1)\.
- \[11\]M\. J\. Daniel\(2020\-10\)Skills aren’t soft or hard — they’re durable or perishable\.Note:Chief Learning OfficerExternal Links:[Link](https://www.chieflearningofficer.com/2020/10/29/skills-arent-soft-or-hard-theyre-durable-or-perishable/)Cited by:[§I](https://arxiv.org/html/2607.14044#S1.p1.1),[§V](https://arxiv.org/html/2607.14044#S5.p1.1)\.
- \[12\]F\. Dell’Acqua, E\. McFowland III, E\. Mollick, H\. Lifshitz, K\. C\. Kellogg, S\. Rajendran, L\. Krayer, F\. Candelon, and K\. R\. Lakhani\(2026\-03\)Navigating the jagged technological frontier: field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality\.Organization Science37\(2\),pp\. 403–423\.External Links:[Document](https://dx.doi.org/10.1287/orsc.2025.21838)Cited by:[§I](https://arxiv.org/html/2607.14044#S1.p2.1)\.
- \[13\]H\. Heickal and A\. Lan\(2026\)A classroom study of LLM\-generated feedback intervention in introductory programming\.Note:arXiv preprint; accepted at IRAISE 2026 \(Festival of Learning\)External Links:2606\.08807,[Document](https://dx.doi.org/10.48550/arXiv.2606.08807)Cited by:[§II\-A](https://arxiv.org/html/2607.14044#S2.SS1.p4.1),[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p4.1)\.
- \[14\]P\. Illanes, S\. Lund, M\. Mourshed, S\. Rutherford, and M\. Tyreman\(2018\-01\)Retraining and reskilling workers in the age of automation\.Technical reportMcKinsey Global Institute\.External Links:[Link](https://www.mckinsey.com/featured-insights/future-of-work/retraining-and-reskilling-workers-in-the-age-of-automation)Cited by:[§I](https://arxiv.org/html/2607.14044#S1.p3.1),[§V](https://arxiv.org/html/2607.14044#S5.p3.1)\.
- \[15\]C\. Isleyet al\.\(2026\)Assessing the quality of AI\-generated exams: a large\-scale field study\.InProc\. AAAI Conf\. Artif\. Intell\.,Vol\.40,pp\. 38626–38634\.External Links:[Document](https://dx.doi.org/10.1609/aaai.v40i45.41205)Cited by:[§III\-E](https://arxiv.org/html/2607.14044#S3.SS5.p1.1)\.
- \[16\]M\. Kapur\(2014\)Productive failure in learning math\.Cognitive Science38\(5\),pp\. 1008–1022\.External Links:[Document](https://dx.doi.org/10.1111/cogs.12107)Cited by:[§III\-D](https://arxiv.org/html/2607.14044#S3.SS4.p1.1)\.
- \[17\]M\. S\. Knowles, E\. F\. Holton, and R\. A\. Swanson\(2015\)The adult learner: the definitive classic in adult education and human resource development\.8th edition,Routledge,London, U\.K\.\.External Links:ISBN 978\-0415739023Cited by:[§II](https://arxiv.org/html/2607.14044#S2.p1.1)\.
- \[18\]A\. LaPrade, J\. Mertens, T\. Moore, and A\. Wright\(2019\-09\)The enterprise guide to closing the skills gap: strategies for building and maintaining a skilled workforce\.Technical reportIBM Institute for Business Value\.External Links:[Link](https://www.ibm.com/thought-leadership/institute-business-value/en-us/report/closing-skills-gap)Cited by:[§I](https://arxiv.org/html/2607.14044#S1.p1.1)\.
- \[19\]H\. Liet al\.\(2025\)ARCHED: a human\-centered framework for transparent, responsible, and collaborative AI assisted instructional design\.InProc\. iRAISE Workshop at AAAI,Proceedings of Machine Learning Research, Vol\.273,pp\. 94–104\.External Links:[Link](https://proceedings.mlr.press/v273/li25a.html)Cited by:[§II\-A](https://arxiv.org/html/2607.14044#S2.SS1.p3.1),[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p2.1)\.
- \[20\]A\. Liu, M\. Sun, L\. Esbenshade, V\. Tian, Z\. Zhang, and K\. He\(2026\)Teacher\-authored prompts for configuring student\-AI dialogue: K12 classroom implementation\.Note:arXiv preprint; journal DOI to be assignedExternal Links:2604\.16738,[Document](https://dx.doi.org/10.48550/arXiv.2604.16738)Cited by:[§II\-A](https://arxiv.org/html/2607.14044#S2.SS1.p4.1)\.
- \[21\]A\. Mishraet al\.\(2024\)Fine\-grained hallucination detection and editing for language models\.InProc\. 1st Conf\. Language Modeling \(COLM\),External Links:[Link](https://openreview.net/forum?id=dJMTn3QOWO)Cited by:[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p3.1),[§III\-C](https://arxiv.org/html/2607.14044#S3.SS3.p1.1)\.
- \[22\]B\. Nguyen, T\. Du, M\. Yu, L\. Angrave, and M\. Jiang\(2025\-07\)QG\-SMS: enhancing test item analysis via student modeling and simulation\.InProc\. 63rd Annu\. Meeting Assoc\. Comput\. Linguistics \(ACL\),Vienna, Austria,pp\. 26152–26168\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.1268)Cited by:[§III\-E](https://arxiv.org/html/2607.14044#S3.SS5.p1.1)\.
- \[23\]T\. N\. Nguyenet al\.\(2026\)Security considerations for multi\-agent AI systems: a scalable design\-based framework and survey\.Note:Manuscript under reviewCited by:[§III\-A](https://arxiv.org/html/2607.14044#S3.SS1.p1.1),[§IV\-B](https://arxiv.org/html/2607.14044#S4.SS2.p1.1),[§VI](https://arxiv.org/html/2607.14044#S6.p1.1)\.
- \[24\]S\. Noy and W\. Zhang\(2023\-07\)Experimental evidence on the productivity effects of generative artificial intelligence\.Science381\(6654\),pp\. 187–192\.External Links:[Document](https://dx.doi.org/10.1126/science.adh2586)Cited by:[§I](https://arxiv.org/html/2607.14044#S1.p2.1)\.
- \[25\]G\. Raulet al\.\(2025\-11\)RAG\-PRISM: a personalized, rapid, and immersive skill mastery framework with adaptive retrieval\-augmented tutoring\.InProc\. IEEE Frontiers in Educ\. Conf\. \(FIE\),Nashville, TN, USA,pp\. 1–9\.External Links:[Document](https://dx.doi.org/10.1109/FIE63693.2025.11328742)Cited by:[§II\-A](https://arxiv.org/html/2607.14044#S2.SS1.p4.1)\.
- \[26\]J\. M\. Reddiget al\.\(2026\)Guidelines for designing AI technologies to support adult learning\.InProc\. ACM Designing Interactive Syst\. Conf\. \(DIS\),Singapore,pp\. 2474–2496\.External Links:[Document](https://dx.doi.org/10.1145/3800645.3813102)Cited by:[§II](https://arxiv.org/html/2607.14044#S2.p1.1)\.
- \[27\]H\. L\. Roediger and J\. D\. Karpicke\(2006\)Test\-enhanced learning: taking memory tests improves long\-term retention\.Psychological Science17\(3\),pp\. 249–255\.External Links:[Document](https://dx.doi.org/10.1111/j.1467-9280.2006.01693.x)Cited by:[§III\-B](https://arxiv.org/html/2607.14044#S3.SS2.p1.1),[§III\-D](https://arxiv.org/html/2607.14044#S3.SS4.p1.1)\.
- \[28\]J\. Savelkaet al\.\(2025\)AI technicians: developing rapid occupational training methods for a competitive AI workforce\.InProc\. 56th ACM Tech\. Symp\. Comput\. Sci\. Educ\. \(SIGCSE TS\),Pittsburgh, PA, USA,pp\. 1029–1035\.External Links:[Document](https://dx.doi.org/10.1145/3641554.3701935)Cited by:[§II\-A](https://arxiv.org/html/2607.14044#S2.SS1.p5.1)\.
- \[29\]A\. Scarlatos, N\. Liu, J\. Lee, R\. Baraniuk, and A\. Lan\(2025\)Training LLM\-based tutors to improve student learning outcomes in dialogues\.InProc\. AIED 2025,Lecture Notes in Computer Science, Vol\.15877,Cham,pp\. 251–266\.External Links:[Document](https://dx.doi.org/10.1007/978-3-031-98414-3%5F18)Cited by:[§III\-D](https://arxiv.org/html/2607.14044#S3.SS4.p1.1)\.
- \[30\]Y\. Song, Y\. Kim, and M\. Iyyer\(2024\-11\)VeriScore: evaluating the factuality of verifiable claims in long\-form text generation\.InFindings Assoc\. Comput\. Linguistics: EMNLP,Miami, Florida, USA,pp\. 9447–9474\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.findings-emnlp.552)Cited by:[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p3.1),[§III\-C](https://arxiv.org/html/2607.14044#S3.SS3.p1.1)\.
- \[31\]D\. R\. Thomas, C\. Borchers, K\. P\. Vanacore, K\. R\. Koedinger, and R\. F\. Kizilcec\(2026\)Modernizing ground truth: four shifts toward improving reliability and validity in AI in education\.InProc\. 27th Int\. Conf\. Artif\. Intell\. Educ\. \(AIED\),Lecture Notes in Computer Science, Vol\.16586,pp\. 117–131\.External Links:[Document](https://dx.doi.org/10.1007/978-3-032-29773-0%5F9)Cited by:[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p5.1)\.
- \[32\]K\. VanLehn\(2011\)The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems\.Educational Psychologist46\(4\),pp\. 197–221\.External Links:[Document](https://dx.doi.org/10.1080/00461520.2011.611369)Cited by:[§III\-D](https://arxiv.org/html/2607.14044#S3.SS4.p1.1)\.
- \[33\]H\. M\. Wallachet al\.\(2025\)Position: evaluating generative AI systems is a social science measurement challenge\.InProc\. 42nd Int\. Conf\. Mach\. Learning \(ICML\),Proceedings of Machine Learning Research, Vol\.267,pp\. 82232–82251\.External Links:[Link](https://proceedings.mlr.press/v267/wallach25a.html)Cited by:[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p5.1),[§III\-C](https://arxiv.org/html/2607.14044#S3.SS3.p1.1)\.
- \[34\]J\. Wang, R\. Xiao, X\. Hou, and J\. Stamper\(2025\)Enabling multi\-agent systems as learning designers: applying learning sciences to AI instructional design\.External Links:2508\.16659,[Link](https://arxiv.org/abs/2508.16659)Cited by:[§II\-A](https://arxiv.org/html/2607.14044#S2.SS1.p3.1),[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p2.1)\.
- \[35\]R\. E\. Wang, A\. T\. Ribeiro, C\. D\. Robinson, S\. Loeb, and D\. Demszky\(2024\)Tutor CoPilot: a human\-AI approach for scaling real\-time expertise\.Note:arXiv preprint; also available as EdWorkingPaper No\. 24\-1054, Annenberg Institute at Brown UniversityExternal Links:2410\.03017,[Document](https://dx.doi.org/10.48550/arXiv.2410.03017)Cited by:[§II\-A](https://arxiv.org/html/2607.14044#S2.SS1.p4.1),[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p4.1)\.
- \[36\]D\. Weitekamp, M\. N\. Siddiqui, and C\. J\. MacLellan\(2025\)TutorGym: a testbed for evaluating AI agents as tutors and students\.InProceedings of the 26th International Conference on Proc\. AIED 2025,pp\. 361–376\.External Links:[Document](https://dx.doi.org/10.1007/978-3-031-98420-4%5F26)Cited by:[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p4.1),[§III\-D](https://arxiv.org/html/2607.14044#S3.SS4.p1.1)\.
- \[37\]World Economic Forum\(2025\-01\)The future of jobs report 2025\.Insight ReportWorld Economic Forum,Geneva, Switzerland\.External Links:[Link](https://www.weforum.org/publications/the-future-of-jobs-report-2025/)Cited by:[§I](https://arxiv.org/html/2607.14044#S1.p1.1)\.
- \[38\]Z\. Xu, X\. Li, Y\. Huan, V\. Minaya, and R\. Yu\(2025\)From course to skill: evaluating large language model performance in curricular analytics\.InProc\. AIED 2025,Lecture Notes in Computer Science, Vol\.15882,Cham,pp\. 203–211\.External Links:[Document](https://dx.doi.org/10.1007/978-3-031-98465-5%5F26)Cited by:[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p5.1)\.
- \[39\]H\. Yao, W\. Xu, J\. Turnau, N\. Kellam, and H\. Wei\(2026\-03\)Instructional agents: reducing teaching faculty workload through multi\-agent instructional design\.InProc\. 19th Conf\. Eur\. Chapter Assoc\. Comput\. Linguistics \(EACL\),Rabat, Morocco,pp\. 4087–4109\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.eacl-long.191)Cited by:[§II\-A](https://arxiv.org/html/2607.14044#S2.SS1.p3.1),[§II\-B](https://arxiv.org/html/2607.14044#S2.SS2.p2.1),[§IV](https://arxiv.org/html/2607.14044#S4.p1.1)\.
- \[40\]Z\. Yaoet al\.\(2025\-04\)MCQG\-SRefine: multiple choice question generation and evaluation with iterative self\-critique, correction, and comparison feedback\.InProc\. NAACL\-HLT,Albuquerque, New Mexico,pp\. 10728–10777\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.naacl-long.538)Cited by:[§III\-E](https://arxiv.org/html/2607.14044#S3.SS5.p1.1)\.
- \[41\]M\. Zhu, L\. Xu, and B\. Ericson\(2025\)A systematic review of research on large language models for computer programming education\.Note:arXiv preprintExternal Links:2506\.21818,[Document](https://dx.doi.org/10.48550/arXiv.2506.21818)Cited by:[§II\-A](https://arxiv.org/html/2607.14044#S2.SS1.p5.1)\.