Evaluating Adaptive Personalization of Educational Readings with Simulated Learners

# Evaluating Adaptive Personalization of Educational Readings with Simulated Learners
Source: [https://arxiv.org/html/2604.16744](https://arxiv.org/html/2604.16744)
Ryan T\. Woo Anmol Rao11footnotemark:1Aryan Keluskar Yinong ChenSchool of Computing and Augmented IntelligenceArizona State University\{rtwoo,arao75,akeluska,yinong\.chen\}@asu\.edu

###### Abstract

We present a framework for evaluating adaptive personalization of educational reading materials with theory\-grounded simulated learners\. The system builds a learning\-objective and knowledge\-component ontology from open textbooks, curates it in a browser\-based Ontology Atlas, labels textbook chunks with ontology entities, and generates aligned reading–assessment pairs\. Simulated readers learn from passages through a Construction–Integration\-inspired memory model withDIME\-style reader factors,KReC\-style misconception revision, and an open New Dale–Chall readability signal\. Answers are produced by score\-based option selection over the learner’s explicit memory state, whileBKTdrives adaptation\. Across three sampled subject ontologies and matched cohorts of 50 simulated learners per condition, adaptive reading significantly improved outcomes in computer science, yielded smaller positive but inconclusive gains in inorganic chemistry, and was neutral to slightly negative in general biology\.

Evaluating Adaptive Personalization of Educational Readings with Simulated Learners

Ryan T\. Woo††thanks:equal contribution\.Anmol Rao11footnotemark:1Aryan Keluskar Yinong ChenSchool of Computing and Augmented IntelligenceArizona State University\{rtwoo,arao75,akeluska,yinong\.chen\}@asu\.edu

## 1Introduction

With more students entering higher education, instructors increasingly face a familiar mismatch: students are expected to learn the same material at the same pace even though their understanding of particular concepts can vary widely\. In an introductory course, one student may already understand a subset of the target concepts, while another may still hold robust misconceptions about those same concepts\. When both students are given identical readings, examples, and explanations, one may be under\-challenged while the other is overwhelmed\. Addressing this in a real classroom typically requires instructors to rewrite materials, manually identify which concepts students struggle with, and provide individualized support\. Those interventions are difficult to scale\.

Intelligent tutoring systems \(ITSs\) aim to address this problem by adapting instruction to the learner\. Recent reviews suggest that ITSs often improve learning and performance, but that their impact depends strongly on the instructional setting, the adaptation mechanism, and the way systems are evaluated\(Huanget al\.,[2025](https://arxiv.org/html/2604.16744#bib.bib8); Létourneauet al\.,[2025](https://arxiv.org/html/2604.16744#bib.bib13)\)\. In many ITS settings, however, the core intervention is an item sequence, a hint policy, or post\-hoc feedback\. Our setting is different: the primary intervention is*reading material*\. Questions are still important, but mainly because they reveal evidence about what the learner knows after reading and provide observations for mastery tracking\.

To evaluate whether adaptive reading policies help, we need simulated learners that can \(i\) learn from text, \(ii\) differ in reading\-related characteristics, \(iii\) preserve misconceptions rather than merely lacking knowledge, and \(iv\) later answer multiple\-choice questions from what they remember rather than from unrestricted access to the source passage\. Simulated learners have long been proposed as tools for teachers, for learners, and for instructional designers conducting formative evaluation\(VanLehnet al\.,[1994](https://arxiv.org/html/2604.16744#bib.bib1)\)\. More recently, a systematic review across AIED\-related fields found uses spanning teacher training, adaptive algorithm development, virtual collaborators, and testing of learning environments, while also concluding that realism and validity are often under\-tested\(Käser and Alexandron,[2024](https://arxiv.org/html/2604.16744#bib.bib16)\)\. Off\-the\-shelf LLM prompting is a poor fit for that requirement because powerful LLMs often answer beyond the intended ability of a partially knowledgeable learner\.Yuanet al\.\([2026](https://arxiv.org/html/2604.16744#bib.bib15)\)characterize this problem as a*competence paradox*and argue for constrained student simulation under an explicit epistemic state specification\.

We therefore present a theory\-grounded simulated reader for adaptive reading personalization\. The system builds a learning\-objective \(LO\) and knowledge\-component \(KC\) ontology from open textbooks, tracks KC mastery withBKT\(Corbett and Anderson,[1994](https://arxiv.org/html/2604.16744#bib.bib4)\), generates ontology\-driven reading and assessment variants, and simulates reading via a comprehension model grounded in the Construction–Integration \(CI\) theory of discourse comprehension\(Kintsch,[1988](https://arxiv.org/html/2604.16744#bib.bib11)\)\. Learner heterogeneity is parameterized withDIME\-style reader factors\(Cromley and Azevedo,[2007](https://arxiv.org/html/2604.16744#bib.bib5); Cromleyet al\.,[2010](https://arxiv.org/html/2604.16744#bib.bib6)\), misconceptions are revised through the Knowledge Revision Components framework \(KReC\)\(Kendeou and O’Brien,[2014](https://arxiv.org/html/2604.16744#bib.bib9); Kendeou,[2024](https://arxiv.org/html/2604.16744#bib.bib10)\), and later answer selection is performed through score\-based option selection over an explicit epistemic boundary\.

Although the broader architecture is designed to integrate with learning\-management systems such as Canvas by Instructure, the present paper focuses on the simulation layer so that the instructional policy can be studied before any full classroom deployment\. This separation allows us to compare adaptive and non\-adaptive reading policies under matched simulated cohorts, while keeping the content pipeline, question format, and mastery tracker fixed\.

Our main contributions are as follows:

- •We formulate adaptive reading personalization as a closed\-loop problem in which reading is the primary intervention and assessment is the observation channel for KC\-level mastery tracking\.
- •We present an ontology\-grounded content pipeline that links open textbook chunks, reading variants, and assessment items to shared LOs and KCs\.
- •We develop a simulated reader that combinesCI,DIME,KReC, and an open New Dale–Chall readability signal to model what a learner comprehends, revises, retains, and later retrieves\.
- •We report multi\-subject experimental results showing that adaptive reading yields domain\-dependent gains: statistically reliable improvements in the sampled computer\-science setting, positive but inconclusive gains in inorganic chemistry, and neutral to slightly negative outcomes in general biology\.

## 2Related Work

### 2\.1Adaptive tutoring and knowledge tracing

Knowledge tracing estimates latent mastery from observed responses, andBKTremains attractive because it is interpretable at the level of individual KCs\(Corbett and Anderson,[1994](https://arxiv.org/html/2604.16744#bib.bib4)\)\. Although deep models can improve prediction, interpretability remains important for deployment and diagnosis\(Ding and Larson,[2021](https://arxiv.org/html/2604.16744#bib.bib7)\)\. We therefore useBKTso that the state driving adaptation remains understandable\.

Most ITS personalization focuses on problem selection, hints, or feedback\. By contrast, we center on explanatory reading material\. Because reading can change knowledge without explicit responses, we separate the learner’s hidden*true*knowledge from the tutor’sBKT\-based*estimate*, and only the latter drives adaptation\.

### 2\.2Simulated learners in educational technology

Simulated learners have long been used in educational technology\.VanLehnet al\.\([1994](https://arxiv.org/html/2604.16744#bib.bib1)\)categorize their use for teachers, learners, and instructional designers; our setting aligns most closely with formative evaluation of instructional policies before deployment\.

Recent work highlights broader uses, including teacher training, hypothesis generation, adaptive algorithm development, virtual collaborators, and environment testing, while also noting that many simulators remain under\-validated for their intended decisions\(Käser and Alexandron,[2024](https://arxiv.org/html/2604.16744#bib.bib16)\)\. This motivates our emphasis on explicit learner state, inspectable memory, and theory\-grounded reading dynamics\.

### 2\.3Reading comprehension and computational reader models

Our model is grounded in the Construction–Integration framework\(Kintsch,[1988](https://arxiv.org/html/2604.16744#bib.bib11)\), where readers form a propositional*textbase*and a richer*situation model*integrating prior knowledge\. This supports modeling partial retention, revision, and distortion of information\.

We extendCIwithDIME, which captures reader differences in background knowledge, vocabulary, and inference\(Cromley and Azevedo,[2007](https://arxiv.org/html/2604.16744#bib.bib5); Cromleyet al\.,[2010](https://arxiv.org/html/2604.16744#bib.bib6)\), andKReC, which models revision of incorrect prior knowledge\(Kendeou and O’Brien,[2014](https://arxiv.org/html/2604.16744#bib.bib9); Kendeou,[2024](https://arxiv.org/html/2604.16744#bib.bib10)\)\. This matters because many systematic wrong answers reflect stable misconceptions rather than random noise\(van den Broek and Kendeou,[2008](https://arxiv.org/html/2604.16744#bib.bib17); Gierlet al\.,[2017](https://arxiv.org/html/2604.16744#bib.bib18)\)\.

Prior work has implementedCI\-like automated readers\(Lemaireet al\.,[2006](https://arxiv.org/html/2604.16744#bib.bib12)\); our contribution is to embed such a reader in a KC\-level adaptive instructional loop\.

### 2\.4Epistemic boundaries and answer modeling

Recent work argues that valid student simulation should expose an explicit epistemic state and bounded competence\(Yuanet al\.,[2026](https://arxiv.org/html/2604.16744#bib.bib15)\)\. We adopt that principle, but do not use free\-form generation for answer choice\. Instead, answers are selected by scoring options from the learner’s memory state, including mastery, misconceptions, retrieved traces, and item difficulty\. This design is efficient, auditable, and consistent with constrained learner behavior\.

### 2\.5Readability and reader–text matching

Reader–text fit is important but should remain simple to operationalize\. We use the New Dale–Chall readability formula\(Chall and Dale,[1995](https://arxiv.org/html/2604.16744#bib.bib3)\), computed viatextstat\(Bansal and Aggarwal,[2026](https://arxiv.org/html/2604.16744#bib.bib2)\), as a reproducible measure of text difficulty combined with learner parameters\.

## 3System Overview

The overall system converts open educational content into a grounded personalization loop\. The pipeline can be summarized in five stages: \(1\) ontology construction and authoring, \(2\) chunk labeling and retrieval indexing, \(3\) ontology\-driven reading and assessment generation, \(4\) theory\-grounded simulated reading and score\-based answer selection, and \(5\)BKT\-based adaptation\.

### 3\.1Ontology Atlas and ontology construction

We first build an ontology of learning objectives and knowledge components from open textbooks\. Each LO maps to one or more KCs\. The ontology is created with an LLM\-assisted workflow that proposes candidate LOs and KCs from source material and then consolidates them into a course\-specific YAML schema\. We use the ontology as a common representational layer for retrieval, generation, mastery tracking, and simulation\.

To make this layer inspectable, we built Ontology Atlas, a browser\-based ontology workspace backed directly by the YAML files used elsewhere in the system\. Atlas exposes a coverage inventory showing chapter, LO, and KC counts; a searchable navigator over IDs, statements, labels, and descriptions; a D3 relationship graph for the chapter→\\rightarrowLO→\\rightarrowKC hierarchy; and a details panel for editing entries, importing/exporting YAML, validating fields, and saving back to disk\. Atlas is therefore not just a visualization layer; it is the authoring interface for the ontology that the generation and experiment services consume\.

In the current codebase, Ontology Atlas surfaces multiple full\-course ontologies, including the three used in the experiments reported here: foundations of computer science \(16 chapters, 53 LOs, 131 KCs\), general biology \(20 chapters, 60 LOs, 172 KCs\), and introductory inorganic chemistry \(12 chapters, 57 LOs, 177 KCs\)\. This coverage view was useful when selecting sampled LOs for the experiments and when checking whether KC granularity was comparable across subjects\.

### 3\.2Chunk labeling and retrieval indexing

The textbook corpus is segmented into fine\-grained chunks and coarser context chunks\. Fine\-grained chunks serve as the primary evidence units for labeling, while coarser chunks provide broader context during downstream generation\. Each fine\-grained chunk is labeled with one or more LOs and KCs using a three\-stage LLM\-as\-judge pipeline:

1. 1\.encode the chunk as a semantic embedding,
2. 2\.retrieve candidate LO and KC labels with nearest\-neighbor search in a vector database, and
3. 3\.prompt an LLM to adjudicate among the candidates and return the final label set\.

The result is a grounded index in which every retained content unit is tied to the ontology and remains linked to its surrounding context\. Because Atlas edits the same YAML that the labeling and generation services read, ontology revisions immediately propagate to retrieval and generation\.

### 3\.3Ontology\-driven reading and assessment generation

Using the labeled corpus, we generate paired readings and assessments that comprehensively cover one LO at a time\. Each generated reading is grounded in retrieved fine\-grained chunks and optional coarser context chunks; generation parameters control explanation depth, example density, misconception refutation, and target difficulty\. Each assessment is a fresh isomorphic multiple\-choice variant aligned to the same LO and KCs rather than an exact reuse of the same item surface form\.

This choice is important for two reasons\. First, exact question reuse would make policy evaluation more sensitive to item wording than to learner state\. Second, our simulated learners do not learn from assessment interactions; only reading updates their hidden semantic state\. Using fresh isomorphic variants therefore preserves comparability across iterations without allowing assessment exposure to become an unintended instructional channel\.

## 4Methodology

### 4\.1Separating hidden learning from observed mastery

The simulated learner maintains two distinct states\. The first is a hidden*true learning state*that governs what the learner actually knows, misunderstands, and remembers\. The second is the tutor\-visibleBKTestimate that is updated only from observed answers\. This separation prevents the simulation from becoming self\-fulfilling: the personalization policy acts on what the tutor*infers*, not on hidden ground truth\.

For learneriiand KCkk, we represent hidden state as

zi=\{mi,k∗,wi,k,1:M,ri,si,ei\},z\_\{i\}=\\\{m^\{\*\}\_\{i,k\},\\,w\_\{i,k,1:M\},\\,r\_\{i\},\\,s\_\{i\},\\,e\_\{i\}\\\},\(1\)wheremi,k∗m^\{\*\}\_\{i,k\}is the strength of correct knowledge for KCkk,wi,k,1:Mw\_\{i,k,1:M\}are strengths of KC\-specific misconceptions,rir\_\{i\}is a reader profile,sis\_\{i\}is a revision\-and\-response profile, andeie\_\{i\}stores recent episodic traces from readings\. The tutor only observes a separate KC mastery estimatem^i,k\\hat\{m\}\_\{i,k\}maintained byBKT\.

Table 1:Learner\-state variables\. Hidden learning and tutor\-visible mastery are kept separate throughout the simulation\.
### 4\.2Reader profiles and readability matching

We do not use vague “learning styles\.” Instead, learner heterogeneity is parameterized through theory\-grounded reader profiles\. FollowingDIME, each learner has parameters for background knowledge, vocabulary, inferencing, and strategy use\(Cromley and Azevedo,[2007](https://arxiv.org/html/2604.16744#bib.bib5); Cromleyet al\.,[2010](https://arxiv.org/html/2604.16744#bib.bib6)\)\. Because the cohort is intended to approximate college freshmen who have completed secondary school, basic decoding is assumed to be near ceiling; the meaningful variability is in comprehension\-relevant factors and in misconception prevalence\.

We combine these learner\-side parameters with a reader–text fit term based on the New Dale–Chall score\. For passagejj, letdjd\_\{j\}be the readability score computed withtextstat\. Letaia\_\{i\}be a scalar readability\-ability parameter for learnerii\. We define

matchi​j=1−\|ai−dj\|6,\\text\{match\}\_\{ij\}=1\-\\frac\{\|a\_\{i\}\-d\_\{j\}\|\}\{6\},\(2\)clipped to\[−1,1\]\[\-1,1\], so higher values indicate a better match between learner and text\. This term does not replace the richer reader profile; it provides a compact difficulty signal that can be integrated into the comprehension update\.

### 4\.3From passages to proposition\-like teaching events

We model reading at the level of comprehension and memory, not time\. Each generated passage is segmented into short teaching events, which serve as proposition\-like units for the simulation\. The current implementation uses a lean operational schema rather than full semantic parsing: each event stores its surface text, linked KC IDs, a clarity score, a refutation\-strength score, and a refresh flag\.

These fields have direct roles in the simulation\.*KC links*determine which latent mastery variables and misconception weights the event can update\.*Clarity*approximates how easy the event is to encode; in the current implementation it is estimated from event length, so shorter events are treated as easier to encode than longer ones\.*Refutation strength*approximates how explicitly the event confronts a likely misconception; in the current implementation it is triggered by lexical cues such asmisconception,incorrectly,rather than,instead of, ordo not, and it scales misconception reduction during learning\.*Refresh*marks lower\-gain reinforcement events when review KCs are included in a passage; operationally, refresh is a lightweight heuristic rather than a discourse parser\.

This representation keeps learner memory close to theCInotion of a textbase while remaining simple enough to generate, score, and inspect at scale\. Encoded events create episodic traces that preserve the event text, KC links, clarity, refutation strength, readability score, and later distortion, making subsequent answer behavior inspectable\. When provenance is needed, it is retained at the reading\-asset level through source chunk identifiers rather than per\-event source spans\.

A passage such as “Plants do not get food from soil\. They make food through photosynthesis\.” would yield an event record like

> \{ proposition\_id: "prop\_001", text: "Plants do not get food from soil\. They make food through photosynthesis\.", kc\_ids: \{"KC\_photosynthesis\_food\_source"\}, clarity: 0\.92, refutation\_strength: 1\.00, is\_refresh: false \}

This is best understood as an operational, proposition\-like teaching event\.

### 4\.4Comprehension as textbase construction and situation\-model integration

The reading update proceeds in two stages\. First, the learner encodes teaching events into a textbase\-like episodic memory\. Second, the learner integrates those events with prior knowledge to update correct knowledge and misconception weights\. Operationally, the encoding strength of eventppfor learneriireading passagejjis modeled as

enci,p=σ\(α\+1\.1​clarityp\+0\.55​matchi​j\+β⊤​ri\+0\.3​\(1−m¯i,p\)\+0\.35attentioni\)\.\\begin\{split\}\\text\{enc\}\_\{i,p\}=\\sigma\\bigl\(&\\alpha\+1\.1\\,\\text\{clarity\}\_\{p\}\+0\.55\\,\\text\{match\}\_\{ij\}\\\\ &\{\}\+\\beta^\{\\top\}r\_\{i\}\+0\.3\(1\-\\bar\{m\}\_\{i,p\}\)\\\\ &\{\}\+0\.35\\,\\text\{attention\}\_\{i\}\\bigr\)\.\\end\{split\}\(3\)Here,m¯i,p\\bar\{m\}\_\{i,p\}is mean prior mastery over the event’s target KCs andβ⊤​ri\\beta^\{\\top\}r\_\{i\}collects vocabulary, inference, strategy, and background\-knowledge effects\. If an event is encoded, it creates an episodic trace whose later activation depends on recency, rehearsal, lexical overlap with the question, and distortion noise\.

Once encoded, an event updates KC state by increasing correct knowledge and reducing misconception weight\. In the implementation, correct\-knowledge gain is stronger for clearer events, better reader–text matches, and learners with higher revision willingness and detail preference\. Misconception reduction is scaled by the event’s refutation strength, consistent withKReC’s emphasis on how prior incorrect knowledge is confronted during reading\(Kendeou and O’Brien,[2014](https://arxiv.org/html/2604.16744#bib.bib9); Kendeou,[2024](https://arxiv.org/html/2604.16744#bib.bib10)\)\. Refresh events receive a smaller gain multiplier than new target events\.

### 4\.5Score\-based answer option selection from learner memory

At assessment time, the learner does*not*receive the original passage\. Instead, the simulator retrieves relevant traces from memory and computes a score for each item from hidden mastery, misconception weight, retrieved\-trace support, and response\-style variables\. The guiding principle is still epistemic boundedness in the sense ofYuanet al\.\([2026](https://arxiv.org/html/2604.16744#bib.bib15)\): the learner answers only from what it has encoded and retained\. However, the current implementation uses score\-based option selection rather than free\-form generative reasoning\.

For itemqq, letm¯i​q\\bar\{m\}\_\{iq\}be mean correct knowledge over the target KCs,w¯i​q\\bar\{w\}\_\{iq\}the mean misconception weight, andτi​q\\tau\_\{iq\}the mean activation of retrieved traces\. The item’s response utility is

ui​q=−bq\+2\.2​m¯i​q\+0\.38​τi​q−1\.25​w¯i​q\+0\.35​attentioni−0\.25​guessi,\\begin\{split\}u\_\{iq\}=\{\}&\-b\_\{q\}\+2\.2\\,\\bar\{m\}\_\{iq\}\+0\.38\\,\\tau\_\{iq\}\\\\ &\{\}\-1\.25\\,\\bar\{w\}\_\{iq\}\+0\.35\\,\\text\{attention\}\_\{i\}\\\\ &\{\}\-0\.25\\,\\text\{guess\}\_\{i\},\\end\{split\}\(4\)wherebqb\_\{q\}is a difficulty offset derived from the item’s difficulty band\. The probability of a correct answer is then

P​\(yi​q=1\)=σ​\(ui​q\),P\(y\_\{iq\}=1\)=\\sigma\(u\_\{iq\}\),\(5\)clipped to a small interior interval to avoid degenerate probabilities\.

If the learner answers incorrectly, the selected distractor is*not*sampled uniformly\. Instead, distractors are scored by lexical overlap between the retrieved trace text and the distractor text/rationale, plus learner\-specific guessing bias\. This makes misconception\-consistent wrong options more likely than random errors when the learner’s memory traces support them\. The simulator also records an explicit epistemic\-state summary for each response, including target KCs, correct\-knowledge support, misconception support, and the IDs of the retrieved traces used during scoring\.

## 5Adaptive Personalization with Bayesian Knowledge Tracing

The tutor maintains a separateBKTmodel for each KC\. After each question response, the tutor updates the KC\-specific mastery estimatem^i,k\\hat\{m\}\_\{i,k\}using standardBKTparameters for initial mastery, learning, guess, and slip\(Corbett and Anderson,[1994](https://arxiv.org/html/2604.16744#bib.bib4)\)\. These estimates drive reading adaptation but do not directly modify hidden learner state\.

The adaptation policy maps low estimated mastery to more supportive readings\. In the adaptive condition, a lowm^i,k\\hat\{m\}\_\{i,k\}can trigger:

- •greater explanatory depth,
- •higher example density,
- •more explicit misconception refutation,
- •repeated coverage of weak KCs within the current LO, and
- •a tighter reader–text match through the target readability signal\.

A higherm^i,k\\hat\{m\}\_\{i,k\}leads to leaner explanations and less redundant scaffolding\. The control condition uses the same ontology, source corpus, and assessment machinery, but the reading variant is held at a fixed default configuration rather than being conditioned onm^i,k\\hat\{m\}\_\{i,k\}\.

## 6Experimental Setup

We ran fixed\-iteration experiments in three subject ontologies: foundations of computer science, general biology, and introductory inorganic chemistry\.[Table 2](https://arxiv.org/html/2604.16744#S6.T2)summarizes the ontology sizes surfaced in Ontology Atlas for those three subjects\.

Table 2:Ontology sizes shown in the Ontology Atlas coverage view for the three ontologies used in the experiments\.For each subject, we sampled four LOs across two chapters, created 50 matched simulated learners per condition, ran three reading–assessment cycles per LO, and generated three KC\-aligned multiple\-choice items per cycle\. This yields 1,800 item\-level responses per condition per subject\.

The adaptive and control runs were matched at initialization: each control run reused the same learner seeds and initial hidden states as its adaptive counterpart\. Assessment delivery context was set tosummative, so quiz answers updated tutor\-visibleBKTstate but did not directly change hidden semantic knowledge\. This keeps reading as the sole instructional intervention in the reported experiments\. We report four outcome families:\(1\) Observed assessment accuracy: mean item accuracy over each condition\.\(2\) Tutor\-sideBKTgain: change in tutor\-visible KC mastery estimates\.\(3\) Hidden mastery gain: descriptive change in hidden correct\-knowledge strength\.\(4\) Misconception reduction: descriptive decrease in hidden misconception weight\.

Because the saved experiment artifacts preserve learner\-level paired trajectories for accuracy andBKTgain, but aggregate hidden mastery and misconception endpoints at the condition level, inferential tests are reported only for the first two outcomes\.

## 7Results

### 7\.1Subject\-level outcomes

[Figure 1](https://arxiv.org/html/2604.16744#S7.F1)and[Table 3](https://arxiv.org/html/2604.16744#S7.T3)summarizes the adaptive\-minus\-control deltas across subjects\. The strongest evidence for adaptive benefit appears in computer science\. Adaptive reading improved mean accuracy from 0\.834 to 0\.865 \(Δ=\+0\.031\\Delta=\+0\.031, 95% CI \[\+0\.004\+0\.004,\+0\.058\+0\.058\],p=0\.026p=0\.026,dz=0\.325d\_\{z\}=0\.325\) and improved learner\-levelBKTgain from 0\.416 to 0\.454 \(Δ=\+0\.038\\Delta=\+0\.038, 95% CI \[\+0\.002\+0\.002,\+0\.074\+0\.074\],p=0\.037p=0\.037,dz=0\.304d\_\{z\}=0\.304\)\.

In inorganic chemistry, adaptive reading also produced positive deltas, but they were smaller and not statistically reliable in this sample: accuracy rose from 0\.834 to 0\.853 \(Δ=\+0\.019\\Delta=\+0\.019, 95% CI \[−0\.004\-0\.004,\+0\.042\+0\.042\],p=0\.104p=0\.104\), whileBKTgain rose from 0\.418 to 0\.427 \(Δ=\+0\.009\\Delta=\+0\.009, 95% CI \[−0\.021\-0\.021,\+0\.040\+0\.040\],p=0\.540p=0\.540\)\. In general biology, adaptive and control were effectively tied on observed performance, with a slight edge to control: accuracy moved from 0\.850 to 0\.845 \(Δ=−0\.005\\Delta=\-0\.005, 95% CI \[−0\.030\-0\.030,\+0\.020\+0\.020\],p=0\.691p=0\.691\) andBKTgain from 0\.383 to 0\.369 \(Δ=−0\.014\\Delta=\-0\.014, 95% CI \[−0\.054\-0\.054,\+0\.025\+0\.025\],p=0\.468p=0\.468\)\.

The lower panel of[Figure 1](https://arxiv.org/html/2604.16744#S7.F1)and[Table 4](https://arxiv.org/html/2604.16744#S7.T4)shows the hidden\-state endpoints descriptively\. Adaptive reading produced larger hidden mastery gains in all three subjects, with deltas of\+0\.048\+0\.048in computer science,\+0\.026\+0\.026in biology, and\+0\.053\+0\.053in inorganic chemistry\. Misconception reduction also favored adaptation in computer science and chemistry, though control was slightly better in biology by0\.0030\.003\.

Table 3:Subject\-level adaptive versus control outcomes\. Accuracy and BKT gain are reported from matched learner\-level analyses\.![Refer to caption](https://arxiv.org/html/2604.16744v1/x1.png)Figure 1:Subject\-level adaptive\-minus\-control deltas\. The upper panel shows paired learner\-level accuracy andBKTgain deltas with 95% confidence intervals\. The lower panel shows descriptive hidden\-gain and misconception\-reduction deltas from the condition summaries\.
### 7\.2Cycle\-level trajectories

[Figure 2](https://arxiv.org/html/2604.16744#S7.F2)shows mean accuracy by cycle for all three subjects\. In computer science, adaptive reading led control in every cycle: 0\.802 vs\. 0\.760 in cycle 1, 0\.890 vs\. 0\.853 in cycle 2, and 0\.904 vs\. 0\.890 in cycle 3\. In inorganic chemistry, adaptive also led in every cycle, with the largest margin in cycle 2 \(0\.868 vs\. 0\.828\)\. In general biology, the two conditions were nearly tied in cycle 1 \(0\.794 vs\. 0\.795\), adaptive moved ahead in cycle 2 \(0\.864 vs\. 0\.849\), and control pulled ahead in cycle 3 \(0\.906 vs\. 0\.877\)\.

This cycle pattern suggests that the adaptive policy is not uniformly better or worse; rather, its effect depends on how the generated readings, item bank, and learner model interact in a given subject\. The biology run is especially informative because it shows that the policy can appear helpful mid\-sequence while still ending slightly behind on observed metrics\.

Table 4:Descriptive hidden\-state outcomes from the saved experiment summaries\. These endpoints are not tested inferentially because the current result format does not preserve learner\-level paired trajectories for them\.![Refer to caption](https://arxiv.org/html/2604.16744v1/x2.png)Figure 2:Mean learner\-level quiz accuracy by fixed iteration cycle for each subject, with 95% confidence intervals\. Adaptive reading shows a clear and consistent advantage in computer science, a smaller but positive advantage in inorganic chemistry, and a mixed pattern in general biology\.

## 8Discussion

The clearest result is that adaptive personalization is promising but domain\-dependent\.[Figure 1](https://arxiv.org/html/2604.16744#S7.F1)shows that the adaptive policy improved both observed accuracy and tutor\-side mastery in the sampled computer\-science setting, and it also produced the largest descriptive gains in hidden knowledge and misconception reduction\. This is the strongest evidence in the current paper that adapting explanation depth, example density, and refutation strength to estimated mastery can improve outcomes when the ontology, generated readings, and assessment items are well aligned\.

The inorganic\-chemistry run is more tentative\. Its positive deltas suggest that the same policy family can generalize beyond computer science, but the confidence intervals remain too wide to claim a reliable effect with the present sample\. The general\-biology run is more cautionary: adaptive reading slightly improved hidden mastery descriptively, yet it did not improve observed accuracy or tutor\-sideBKTgain\. That mismatch matters because it suggests that a personalization policy can move the latent learner state in a favorable direction without that change being fully visible through the current assessment and scoring layer\.

This divergence between hidden and observed outcomes is one of the main takeaways from the results\. In the current simulator, answers depend not only on hidden KC knowledge but also on retrieved episodic traces, misconception weight, item difficulty, attention stability, and distractor overlap\. A learner can therefore end a reading cycle with somewhat higher hidden knowledge but still perform similarly on observed items if the item bank is noisy, if retrieval is brittle, or if distractors remain highly competitive\.

The cycle\-level results in[Figure 2](https://arxiv.org/html/2604.16744#S7.F2)reinforce the same point\. Computer science shows a stable adaptive advantage from cycle 1 onward, suggesting that the reading variants are helping early and consistently\. In contrast, biology shows a crossover pattern: adaptive is slightly ahead mid\-sequence but behind by the final cycle\. That kind of crossover is exactly the kind of behavior that an offline simulator should surface before live deployment\.

The results also clarify why the Ontology Atlas is useful\. Adaptive benefit varied by LO and by subject, which means ontology quality and KC granularity are not merely upstream bookkeeping issues\. If an LO is too broad, if its KCs are diffuse, or if retrieved evidence does not align tightly with the target KCs, both personalization and evaluation become harder to interpret\. Atlas helped us inspect ontology scale, sample comparable objectives across subjects, and revise the LO/KC hierarchy without changing the downstream code paths\.

## 9Conclusion

We presented a full\-stack framework for evaluating adaptive personalization of educational readings with theory\-grounded simulated learners\. The framework links ontology\-grounded content generation, KC\-levelBKTmastery tracking, a browser\-based Ontology Atlas for ontology authoring,CI\-based comprehension,DIME\-style reader heterogeneity,KReC\-style misconception revision, and score\-based answer selection over an explicit learner memory state\. Across three sampled subject ontologies, adaptive reading yielded statistically reliable gains in the computer\-science condition, smaller positive but inconclusive gains in inorganic chemistry, and neutral to slightly negative observed outcomes in general biology while still improving hidden mastery descriptively\. We view this as a practical bridge between NLP\-based content generation and cognitively plausible learner simulation for educational personalization, and as an offline testbed for screening adaptive reading policies before live classroom deployment\.

## Limitations

This work has several limitations\. First, simulated learners are not substitutes for real students\. Even with explicit misconceptions and inspectable memory, a simulated reader remains a model of plausible learning dynamics rather than a direct measurement of human learning\. This caution is consistent with the validity gap highlighted byKäser and Alexandron \([2024](https://arxiv.org/html/2604.16744#bib.bib16)\)\. Second, the current design focuses on comprehension and memory, not reading time, attention shifts during reading, or fatigue over long sessions\.

Third, the present implementation uses a simplified score\-based answer model rather than a richer generative answerer\. This keeps runs fast and reproducible, but it may under\-represent some reasoning phenomena that would matter in more open\-ended assessments\. Fourth, the quality of the ontology and proposition\-like teaching\-event extraction directly affects the quality of both generation and simulation; errors in LO/KC labeling can propagate downstream\. Fifth, the New Dale–Chall score is intentionally simple and open, but it is still only one aspect of reader–text fit\. Finally, because the saved result format aggregates hidden\-state endpoints at the condition level, inferential testing is currently limited to observed accuracy and tutor\-sideBKToutcomes\.

### Ethics Statement

The system described here uses open educational content and simulated rather than human learners\. The main ethical motivation for the simulation framework is to reduce the cost of evaluating personalization policies before classroom deployment\. Even so, simulation\-based evidence should not be overstated: decisions that materially affect real students should ultimately be validated with human\-subject studies under appropriate institutional review and privacy safeguards\.

## References

- S\. Bansal and C\. Aggarwal \(2026\)textstat: calculate statistical features from text\.Note:Python package, version 0\.7\.13Accessed 2026\-03\-30External Links:[Link](https://pypi.org/project/textstat/)Cited by:[§2\.5](https://arxiv.org/html/2604.16744#S2.SS5.p1.1)\.
- J\. S\. Chall and E\. Dale \(1995\)Readability revisited: the new dale–chall readability formula\.Brookline Books,Cambridge, MA\.External Links:ISBN 9781571290083Cited by:[§2\.5](https://arxiv.org/html/2604.16744#S2.SS5.p1.1)\.
- 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:[§1](https://arxiv.org/html/2604.16744#S1.p4.1),[§2\.1](https://arxiv.org/html/2604.16744#S2.SS1.p1.1),[§5](https://arxiv.org/html/2604.16744#S5.p1.1)\.
- J\. G\. Cromley and R\. Azevedo \(2007\)Testing and refining the direct and inferential mediation model of reading comprehension\.Journal of Educational Psychology99\(2\),pp\. 311–325\.External Links:[Document](https://dx.doi.org/10.1037/0022-0663.99.2.311)Cited by:[§1](https://arxiv.org/html/2604.16744#S1.p4.1),[§2\.3](https://arxiv.org/html/2604.16744#S2.SS3.p2.1),[§4\.2](https://arxiv.org/html/2604.16744#S4.SS2.p1.1)\.
- J\. G\. Cromley, L\. E\. Snyder\-Hogan, and U\. A\. Luciw\-Dubas \(2010\)Reading comprehension of scientific text: a domain\-specific test of the direct and inferential mediation model of reading comprehension\.Journal of Educational Psychology102\(3\),pp\. 687–700\.External Links:[Document](https://dx.doi.org/10.1037/a0019452)Cited by:[§1](https://arxiv.org/html/2604.16744#S1.p4.1),[§2\.3](https://arxiv.org/html/2604.16744#S2.SS3.p2.1),[§4\.2](https://arxiv.org/html/2604.16744#S4.SS2.p1.1)\.
- X\. Ding and E\. C\. Larson \(2021\)On the interpretability of deep learning based models for knowledge tracing\.External Links:2101\.11335,[Link](https://arxiv.org/abs/2101.11335)Cited by:[§2\.1](https://arxiv.org/html/2604.16744#S2.SS1.p1.1)\.
- M\. J\. Gierl, O\. Bulut, Q\. Guo, and X\. Zhang \(2017\)Developing, analyzing, and using distractors for multiple\-choice tests in education: a comprehensive review\.Review of Educational Research87\(6\),pp\. 1082–1116\.External Links:[Document](https://dx.doi.org/10.3102/0034654317726529)Cited by:[§2\.3](https://arxiv.org/html/2604.16744#S2.SS3.p2.1)\.
- X\. Huang, W\. Xu, and R\. Liu \(2025\)Effects of intelligent tutoring systems on educational outcomes: evidence from a comprehensive analysis\.International Journal of Distance Education Technologies23\(1\),pp\. 1–25\.External Links:[Document](https://dx.doi.org/10.4018/IJDET.368420)Cited by:[§1](https://arxiv.org/html/2604.16744#S1.p2.1)\.
- T\. Käser and G\. Alexandron \(2024\)Simulated learners in educational technology: a systematic literature review and a turing\-like test\.International Journal of Artificial Intelligence in Education34\(2\),pp\. 545–585\.External Links:[Document](https://dx.doi.org/10.1007/s40593-023-00337-2)Cited by:[§1](https://arxiv.org/html/2604.16744#S1.p3.1),[§2\.2](https://arxiv.org/html/2604.16744#S2.SS2.p2.1),[Limitations](https://arxiv.org/html/2604.16744#Sx1.p1.1)\.
- P\. Kendeou and E\. J\. O’Brien \(2014\)The knowledge revision components \(krec\) framework: processes and mechanisms\.InProcessing Inaccurate Information: Theoretical and Applied Perspectives from Cognitive Science and the Educational Sciences,D\. N\. Rapp and J\. L\. G\. Braasch \(Eds\.\),pp\. 353–377\.Cited by:[§1](https://arxiv.org/html/2604.16744#S1.p4.1),[§2\.3](https://arxiv.org/html/2604.16744#S2.SS3.p2.1),[§4\.4](https://arxiv.org/html/2604.16744#S4.SS4.p2.1)\.
- P\. Kendeou \(2024\)A theory of knowledge revision: the development of the krec framework\.Educational Psychology Review36\(2\),pp\. 44\.External Links:[Document](https://dx.doi.org/10.1007/s10648-024-09885-y)Cited by:[§1](https://arxiv.org/html/2604.16744#S1.p4.1),[§2\.3](https://arxiv.org/html/2604.16744#S2.SS3.p2.1),[§4\.4](https://arxiv.org/html/2604.16744#S4.SS4.p2.1)\.
- W\. Kintsch \(1988\)The role of knowledge in discourse comprehension: a construction\-integration model\.Psychological Review95\(2\),pp\. 163–182\.External Links:[Document](https://dx.doi.org/10.1037/0033-295X.95.2.163)Cited by:[§1](https://arxiv.org/html/2604.16744#S1.p4.1),[§2\.3](https://arxiv.org/html/2604.16744#S2.SS3.p1.1)\.
- B\. Lemaire, G\. Denhière, C\. Bellissens, and S\. Jhean\-Larose \(2006\)A computational model for simulating text comprehension\.Behavior Research Methods38\(4\),pp\. 628–637\.External Links:[Document](https://dx.doi.org/10.3758/BF03193895)Cited by:[§2\.3](https://arxiv.org/html/2604.16744#S2.SS3.p3.1)\.
- A\. Létourneau, M\. Deslandes Martineau, P\. Charland, J\. A\. Karran, J\. Boasen, and P\. Léger \(2025\)A systematic review of AI\-driven intelligent tutoring systems \(ITS\) in K\-12 education\.npj Science of Learning10\(1\),pp\. 29\.External Links:[Document](https://dx.doi.org/10.1038/s41539-025-00320-7)Cited by:[§1](https://arxiv.org/html/2604.16744#S1.p2.1)\.
- P\. van den Broek and P\. Kendeou \(2008\)Cognitive processes in comprehension of science texts: the role of co\-activation in confronting misconceptions\.Applied Cognitive Psychology22\(3\),pp\. 335–351\.External Links:[Document](https://dx.doi.org/10.1002/acp.1418)Cited by:[§2\.3](https://arxiv.org/html/2604.16744#S2.SS3.p2.1)\.
- K\. VanLehn, S\. Ohlsson, and R\. Nason \(1994\)Applications of simulated students: an exploration\.Journal of Artificial Intelligence in Education5\(2\),pp\. 135–175\.Cited by:[§1](https://arxiv.org/html/2604.16744#S1.p3.1),[§2\.2](https://arxiv.org/html/2604.16744#S2.SS2.p1.1)\.
- Z\. Yuan, Y\. Xiao, M\. Li, W\. Xuan, R\. Tong, M\. T\. Diab, and T\. Mitchell \(2026\)Towards valid student simulation with large language models\.External Links:2601\.05473,[Document](https://dx.doi.org/10.48550/arXiv.2601.05473),[Link](https://arxiv.org/abs/2601.05473)Cited by:[§1](https://arxiv.org/html/2604.16744#S1.p3.1),[§2\.4](https://arxiv.org/html/2604.16744#S2.SS4.p1.1),[§4\.5](https://arxiv.org/html/2604.16744#S4.SS5.p1.1)\.