AIEd Knowledge BaseAIEd KB
On this page

Concept

Large Language Models (LLMs)

created 2026-08-09 · updated 2026-11-05

Large Language Models (LLMs) — neural network models trained on vast text corpora that generate human-like text, powering most modern AI in education applications. LLMs are the computational backbone of generative AI tutoring, assessment, and content generation in education.

Questions to Consider

  • What do you believe an AI chatbot 'knows' when it answers you? The page frames LLMs as generating probable text rather than retrieving verified facts — how does that distinction change how much you would trust a model's explanations?
  • LLMs are described as the engine behind most modern AI education tools — tutoring, grading, content generation, and even diagnosing what students know. Of these uses, which do you think is most and least appropriate for a probabilistic text generator, and why?
  • The page reports that three different LLMs produced sharply divergent support plans for the same learning-analytics input, each with different demographic assumptions. If models aren't interchangeable as advisors, what does that mean for an institution that adopts one?
  • Because LLM output is sensitive to prompts and settings, two people can get very different results from the same model. How should this influence how you — as a learner or designer — phrase requests, and how much you trust a single output?
  • A key limitation is hallucination — plausible-sounding but ungrounded content. In a tutoring or grading context, what would it take for you to feel confident the model wasn't inventing something, and what safeguards would you demand before letting it assess a real student?

Introduction

LLMs as the engine of AIED

LLMs are the most-referenced concept in the knowledge base (60+ articles) because they underpin nearly every AI education application:

  • Tutoring: AI tutors use LLMs for dialogue, explanation, and Problem Solving guidance. Training and fine-tuning adapts general LLMs for educational use.

  • Assessment: Grading systems, essay scoring, and item difficulty prediction leverage LLM capabilities. Razavi and Powers (2026) show GPT-4o can estimate the difficulty of K-5 math and reading items (N = 5170) calibrated under the Rasch IRT model: zero-shot ratings correlated moderately-to-strongly with true difficulties (r = 0.83 math, r = 0.81 reading) but varied by grade, while a feature-based strategy in which the LLM extracts cognitive and linguistic features for tree-based models reached correlations up to r = 0.87 — evidence that structured feature extraction can outperform a single holistic LLM judgment. Across the aggregate grading literature, a PRISMA-guided systematic review of 42 empirical studies (2023–2025) concludes that LLMs match human raters on short, well-structured tasks with detailed rubrics yet cannot fully replace human judgment on complex, open-ended, or subjective work, and that model version is a dominant determinant of grading quality (Can ChatGPT Replace the Teacher in Assessment? A Review of Research on the Use of Large Language Models in Grading and Providing Feedback). Reliability also varies sharply by item type: Falahat et al. (2026) found ChatGPT-5 matched faculty closely on objective pharmacy-exam items (CCC 0.935–1.000) but was unreliable on short-answer (CCC ≈0) and essay (0.341–0.854) items, and that providing a rubric did not consistently improve agreement.

  • Generated item labels track surface form, not difficulty. A 378-item audit found an LLM's Easy/Medium/Hard labels rose in lockstep with its own co-generated Bloom level (ρ=0.90) and with stem length (15.9 to 30.2 words) yet correlated with empirical item difficulty at only ρ=0.06 over 7,888 student responses, evidence that generation-time difficulty metadata describes formatting rather than demand on prior knowledge (An & Wang (2026)).

  • Generated items can still match expert items. o3-mini in a generate→judge→revise loop built exams for 71 classes; under a Bayesian hierarchical 2PL IRT model the AI items were easier but more discriminating (ᾱ = 1.3 vs. 1.2) and more informative (I_max = 3.85 vs. 2.61) than expert AP Statistics items (Isley et al. (2025)).

  • Annotation quality is a design outcome, not a single-prompt property. Bernado et al. (2026) prompted a five-model panel to judge human-readable assertions about each utterance instead of emitting a label directly, and a transparent classifier over those binary judgments (macro-F1 0.673, Cohen's κ 0.688) beat the best published direct prompting of a frontier model (0.61 macro-F1) while trailing a fine-tuned RoBERTa-base encoder (0.76); cross-model agreement functioned as a screening device, not as validity evidence.

  • Multimodal AI reasoning LLMs as graders: when a multimodal, reasoning-capable LLM (GPT-o4-mini) graded a 296-student handwritten general-chemistry exam page-by-page against rubric images, single-run total scores were highly reproducible (ICC(A,1) = 0.967; averaging five runs reached 0.993) and agreed strongly with TA totals (R² = 0.91), yet item-level reliability was sharply format-dependent — textual and reaction-equation answers graded well while drawing and graphing were worse than random (background grids distract AI vision). This shows an LLM grader's trustworthiness is a function of response format and task, not just raw model capability, and that selective deferral via confidence filters is needed for high-stakes use (Assisting the grading of a handwritten general chemistry exam with artificial intelligence).

  • Content: Generative AI content creation relies on LLMs. Question generation and video generation are LLM-driven.

  • Safety: Pedagogical Safety, Hallucination Risk, and SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems research examine LLM-specific risks.

  • Diagnosis: Knowledge Tracing and Cognitive Diagnosis increasingly incorporate LLMs for richer student modeling. Grounding matters enormously for error diagnosis: Reddig, Arora & MacLellan (2025) showed that supplying GPT-4 the tutor interface structure plus Bayesian Knowledge Tracing skill estimates raised logical-error identification from 40% to 81% on factoring (overall error diagnosis ~87.8%), while multi-step problems and responses with several errors remained weak cases and hallucinated "common-misconception" diagnoses persisted — evidence that an LLM's diagnostic value is as much a function of the structured context and learner-model signals it receives as of the model itself.

  • Verification separated from generation — with correlated failure modes. Hossain et al. (2026) restrict tutor retrieval to instructor-approved course material and route claims through an architecturally separate DeBERTa-v3-large-MNLI verifier, but caution both models share broad web training and can fail in correlated ways, and the verifier cannot check a code trace without executing it.

  • Assessment model shift (2017–2024): Morley et al.'s scoping review of auto-marking short-answer science questions traces the field's move from fine-tuning smaller BERT models (dominant through 2021) toward prompting larger LLMs (GPT-1/2/3.5/4) from roughly 2022 — adopted via Prompt Engineering rather than fine-tuning — with domain-augmented models, rubric-aware prompting, and chain-of-thought lifting accuracy. Yet GPT models were rarely benchmarked against BERT on standard corpora, few auto-markers could explain their marks, and bias was seldom examined, cautions that apply to LLM assessment generally (Auto-marking short answer questions in science: The foundational years of transformer-based models from BERT to GPT-4).

Model-specific research

The knowledge base covers both general-purpose LLMs (GPT-4, Claude) and education-specific adaptations. Small language model benchmarks compare SLM performance for tutoring, and Suvernev et al. (2026) argue SLMs are architecturally appropriate for agentic tutoring when deployed as tool-augmented planners — deterministic tools carry the load a small model would otherwise handle implicitly — with pedagogical control residing in the escalation policy rather than the model. Educational alignment research addresses how to make LLMs pedagogically appropriate. A classroom study across three frontier families — Oppenheimer, Cash & Connell Pensky (2025) — found that ChatGPT, Gemini, or Claude could act as collaborative critique partners for argumentative writing: over a semester of iterative essays, students improved on argument quality, prompt engineering, and response-to-AI feedback by roughly a full standard deviation each (all p < .001) and engaged deeply (87.8% rebutting LLM claims), positioning general-purpose LLMs as viable collaborative learning partners rather than mere answer generators.

A complementary line of work reframes LLMs from static graders into emulators of pedagogical reasoning. Yaşar et al. (2026) showed that GPT-4, scaffolded with a semantically precise, iteratively co-refined rubric, could approximate human evaluative judgment in design-based learning: initial LLM–human agreement was poor (Cronbach's Alpha = 0.393; Kappa −0.06 to 0.18), but iterative rubric refinement raised mean agreement from 54.75% to 81.25% (final Alpha = 0.798, Kappa 0.40–0.55), and K-means clustering of human and LLM score matrices showed highly correlated centroids (r = 0.89). The study positions the rubric as a mediating interface between human pedagogical intent and machine inference — evidence that off-the-shelf LLMs are not interchangeable as evaluators either, and that their assessment behavior is a design outcome shaped by the rubric and prompts they are given. Raw model capability differentiates grading too: benchmarking eleven GenAI and sentence-embedding models on 1,885 open-ended responses, Pecuchova, Benko & Drlik (2025) found only GPTo1 reached almost-perfect agreement with expert human graders (Fleiss' Kappa 0.82), with Claude3 and PaLM2 slightly behind, while reference-aligned models such as BERT fell far short — showing that frontier-model context-sensitivity matters for reliable open-ended assessment. Model differences also matter for high-stakes downstream uses. López-Pernas et al. (2026) showed that three LLMs produced sharply divergent student-support prescriptions for the same learning-analytic input, and each imposed distinct demographic priors on the learner profiles they generated — evidence that off-the-shelf LLMs are not interchangeable as prescriptive advisors. Similarly, Olvet et al. (2026) found that GPT-4's scoring of pre-clerkship medical open-ended questions climbed to substantial-to-almost-perfect inter-rater agreement with faculty (weighted kappa up to 0.94) only after three rounds of iterative rubric refinement, and fell to moderate (κw = 0.54) on a holistic-rubric item — reinforcing that rubric design, not raw capability alone, is the decisive lever for LLM scoring reliability. Model-specific behavior also shows up in how LLMs respond to skeptical users: an algorithmic audit queried ten frontier LLMs 500 times each with a rural-Montana K-12 AI-skeptic persona to test whether AI systems consulted by skeptical users are predisposed to encourage adoption. Eight of ten acknowledged user concerns then redirected to AI-engagement framings; composite scores spanned from 3.85 (Claude Sonnet) to 7.52 (Gemini 3.1 Pro Preview), with a cross-family AI scorer panel clearing Cohen's kappa >= 0.70. The pattern was a model-dependent design outcome. Model capability also depends on how models are combined: Bird (2026) fine-tuned eight state-of-the-art transformers (BERT, ELECTRA, RoBERTa, XLNet, ERNIE, ALBERT, DistilBERT, Longformer) to classify English literature by UK Key Stage, finding the best unimodal transformer (BERT) reached only an F1 of 0.75 — while fusing a fine-tuned ELECTRA with a computational-linguistics neural network lifted F1 to 0.996, showing that transformer text classification alone is limited and that fusion with complementary features is where the gains lie.

LLMs also affirm preferentially: across 11 models, AI responses affirmed users 49% more than human responses, and more sycophantic replies drew higher ratings, raising trust and continued use — a failure mode for coaching, tutoring or feedback, where constructive challenge is the point (Potel & Kumashiro (2026)).

NER coding evidence shows task-dependent superiority rather than a single best model. On a 600-token human-consensus named-entity benchmark, Wang, Pritchard & Rosenberg (2026) found GPT-5.5 led seven-type coding of the 41 gold-entity rows (accuracy 0.805, κ = .726) against spaCy (0.634, κ = .488), while spaCy won binary detection F1 0.734 versus GPT-5.5's 0.654 — model choice follows the downstream analytic task, not any universal best.

Common Misconceptions

  • "The model's raw capability determines how well it can diagnose errors." Grounding does: Reddig, Arora and MacLellan (2025) showed supplying GPT-4 the tutor interface substantially changed its error diagnosis. What the model is given about the learner's context does more work than which model it is.
  • "Any frontier model behaves the same as a learning partner." Oppenheimer, Cash and Connell Pensky (2025) studied three frontier families — ChatGPT, Gemini and Claude — in a classroom comparison, precisely because the families are not interchangeable as collaborators.
  • "Better models mean less Scaffolding needed." GPT-4 scaffolded with a semantically precise, iteratively co-refined rubric could approximate human evaluative judgment in design-based research. The approximation came from the rubric refined over iterations, not from the model alone.

Connected Concepts

Connected Articles

Connected FAQs

Metadata

Technologies and techniquesGenerative AI, Intelligent Tutoring, Prompt Engineering, RAG (Retrieval-Augmented Generation)
Assessment and measurementAutomated Assessment
Equity, ethics, and responsible useHallucination Risk, Pedagogical Safety
Foundations of AI in educationAI Literacy
SourceHistoryEditChat

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.