Three-Layer Instructional Framework

• Three-Layer Instructional Framework is a design approach that decomposes education into three distinct, interdependent layers: knowledge representation, adaptive planning, and evaluative feedback.
• It enables targeted control over instructional design by separating content diagnosis, pedagogical transformation, and outcome verification, each with specific operational roles.
• Empirical studies demonstrate that this layered approach enhances pedagogical alignment and instructional effectiveness across diverse educational and AI-assisted applications.

Across the cited literature, the term Three-Layer Instructional Framework denotes a class of architectures that decompose instructional design, tutoring, educational content generation, or discovery-oriented reasoning into three linked strata, each with a distinct epistemic and operational role. In recent work, these strata have been instantiated as Knowledge–Learning–Instruction (KLI) in multi-agent lesson design, Generation–Evaluation/Alignment–Human Orchestration/Decision in AI-assisted objective authoring, Pedagogical State Modeling–Script-Guided Structured Control–Learning-Oriented Evaluation Metrics in instructional video generation, Perception–Orchestration–Elicitation in Socratic tutoring evaluation, instructional goals–instructional processes–instructional materials in ontology-based educational technology, and Search and Retrieval–Model Formation via Qualitative Reasoning–Execution, Optimization, Refinement in scientific discovery (Wang et al., 20 Aug 2025, Li et al., 11 Mar 2025, Wu et al., 10 May 2026, Liu et al., 8 Aug 2025, Chimalakonda et al., 2018, Liao, 11 Jun 2026). A separate game-based synthesis explicitly consolidates four layers into a three-layer micro–macro–meta framework while preserving shared feedback dynamics across all layers (Beatty, 2014).

1. Variants and definitional range

The frameworks differ in immediate purpose, but each paper assigns a non-interchangeable function to each layer. In KLI, the layers correspond directly to Knowledge Components (KCs), Learning Processes, and Instructional Principles (Wang et al., 20 Aug 2025). In ARCHED, the layers are Generation (LOGS), Evaluation/Alignment (OAE), and Human Orchestration/Decision (Li et al., 11 Mar 2025). In EduStory, the layers are Pedagogical State Modeling, Script-Guided Structured Control, and Learning-Oriented Evaluation Metrics (Wu et al., 10 May 2026). GuideEval uses Perception, Orchestration, and Elicitation as a behavioral framework for tutoring (Liu et al., 8 Aug 2025). IDont separates instructional goals, instructional processes, and instructional materials (Chimalakonda et al., 2018). In AI for scientific discovery, the three layers are Search and Retrieval, Model Formation via Qualitative Reasoning, and Execution, Optimization, Refinement (Liao, 11 Jun 2026).

Framework	Layer 1	Layer 2	Layer 3
EduStory	Pedagogical State Modeling	Script-Guided Structured Control	Learning-Oriented Evaluation Metrics
KLI in MAS	Knowledge Components	Learning Processes	Instructional Principles
ARCHED	Generation (LOGS)	Evaluation/Alignment (OAE)	Human Orchestration/Decision
GuideEval	Perception	Orchestration	Elicitation
IDont	instructional goals	instructional processes	instructional materials
AI in scientific discovery	Search and Retrieval	Model Formation via Qualitative Reasoning	Execution, Optimization, Refinement

This variation is substantive rather than terminological. Some frameworks are aimed at instructional design for teachers and curriculum developers, some at interactive tutoring, some at multimodal generation, and some at scientific reasoning. A plausible implication is that “three-layer” here names a recurrent design logic rather than a single canonical ontology.

2. Layer 1: representation of knowledge, state, and learner condition

In the cited frameworks, the first layer typically specifies what is being taught, what state the system believes it is in, or what learner condition must be diagnosed before any adaptation occurs. EduStory formalizes this most explicitly by modeling each shot $t$ with a pedagogical state

$S_t = \bigl(E_t,\; R_t,\; C\bigr),$

where $E_t$ is the set of knowledge entities introduced or active at shot $t$, $R_t \subseteq E_t \times E_t \times \mathcal{L}$ is a typed relation graph with $$\mathcal{L} = \{\mathrm{causes},\mathrm{quantifies},\mathrm{derives},\mathrm{instantiates}\}$$, and $C$ is a global set of domain-specific constraints such as equation balance, unit consistency, and directional conventions. State transitions are driven by pedagogical actions $a_t \in A$, with $$A = \{\mathrm{Introduce},\mathrm{Derive},\mathrm{Apply},\mathrm{Summarize}\}$$, yielding

$\delta\!\bigl(S_t,\; a_t\bigr) = S_{t+1}.$

The same layer also tracks concept dependency DAGs embedded in $S_t = \bigl(E_t,\; R_t,\; C\bigr),$0 and temporal logic constraints that forbid contradiction or mutation across shots (Wu et al., 10 May 2026).

KLI expresses the first layer as Knowledge Components, defined as facts, concepts, principles, procedures, and skills that learners should acquire. The framework’s premise is that KC types determine appropriate learning processes and, downstream, instructional principles. In the systems described in the paper, the KC Agent enumerates target KCs given subject, grade, standards, and target objective, and outputs a KC list with types (Wang et al., 20 Aug 2025).

IDont represents the first layer through GoalsOntology. Its central class is LearningGoal, which is classified using revised Bloom’s taxonomy and the ABCD model, and enriched with properties for priority, prerequisites, progress, evaluation, and realization via instructional processes and materials. Goal granularity is explicitly aligned to process granularity through PlayLevel, ActLevel, SceneLevel, InstructionLevel, and goals are linked to other layers by object properties such as isAchievedByProcess, usesContent, hasEvaluation, and runsInEnvironment (Chimalakonda et al., 2018).

GuideEval recasts the first layer as Perception, where the tutor must infer one of four operational states: Accurate, Erroneous, Comprehension, or Confusion. The layer is scored by P-Affirm and P-Redirect. For Accurate or Comprehension, the tutor should explicitly affirm correctness and understanding; for Erroneous or Confusion, it should explicitly identify the error or misconception and signal the need for correction. The paper emphasizes that “need for redirection” arises whenever Erroneous or Confusion states are detected (Liu et al., 8 Aug 2025).

ARCHED’s first layer is generative rather than diagnostic. LOGS takes educator-defined parameters such as grade level, subject area, and desired Bloom’s taxonomy levels, and produces multiple candidate learning objectives. The paper specifies that objectives must conform to educator-specified Bloom levels and contextual parameters, and are designed to be measurable, clear, and technically accurate for the target discipline (Li et al., 11 Mar 2025).

Taken together, these first layers show that a three-layer framework often begins with an explicit representational commitment: a state machine, a KC map, a goal ontology, a learner-state diagnosis, or a parameterized objective set. This suggests that the first layer is the principal site where instructional assumptions become formal objects.

3. Layer 2: transformation, planning, alignment, and model formation

The middle layer typically governs how a representation is transformed into an adaptive instructional move, a structured plan, or a revised conceptual model. In EduStory, Script-Guided Structured Control maps a plain-text lesson description $S_t = \bigl(E_t,\; R_t,\; C\bigr),$1 to a two-level structured plan,

$S_t = \bigl(E_t,\; R_t,\; C\bigr),$2

where phases $S_t = \bigl(E_t,\; R_t,\; C\bigr),$3 cover a canonical sequence—Introduction, Explanation, Application, Summary—and each phase holds shot-level pedagogical actions with expected entities and constraint identifiers drawn from $S_t = \bigl(E_t,\; R_t,\; C\bigr),$4. The planner is implemented with a prompted LLM that emits structured JSON control tokens per shot, and generation is conditioned on the current state:

$S_t = \bigl(E_t,\; R_t,\; C\bigr),$5

This layer therefore organizes long-horizon generation by state-conditioned prompts rather than a single long prompt (Wu et al., 10 May 2026).

KLI distributes its middle-layer work across Learning Processes and, in the implemented systems, associated agent roles. The Learning Process Agent selects processes such as memory/fluency building, induction and refinement, and understanding/sense-making, depending on the KC type and target cognitive demand. In the sequential MAS-Roles system, this is followed by an Instructional Principle Agent that chooses teaching tactics aligned to those processes, such as spacing and testing, worked examples, and prompted self-explanation. In MAS-CMD, persona agents independently generate KLI-guided drafts and then refine them through multi-turn peer critique before a Decision Agent selects the final product (Wang et al., 20 Aug 2025).

IDont’s middle layer is ProcessOntology, which captures executable instructional designs through the hierarchy play–act–scene–instruction. The ontology includes InstructionalDesignModel, FirstPrinciples such as ActivationPrinciple and ApplicationPrinciple, activity types such as LearningActivity, SupportActivity, GuidanceActivity, and ReflectionActivity, and granular guideline classes such as PlayGuidelines, ActGuidelines, SceneGuidelines, and InstructionGuidelines. The framework’s ProcessPattern is designed to standardize process design while allowing acts and scenes to be replaced or recombined (Chimalakonda et al., 2018).

In ARCHED, the second layer is Evaluation/Alignment (OAE) rather than process execution. OAE evaluates generated or imported objectives for Bloom-level classification, distribution across cognitive levels, and suggested improvements. The framework explicitly separates evaluation from generation “to maintain accountability and transparency,” and uses step-by-step reasoning chains to make the analysis traceable (Li et al., 11 Mar 2025).

GuideEval’s second layer is Orchestration, which selects pedagogical strategies aligned to the inferred learner state. It is scored by O-Advance for Accurate or Comprehension and O-Reconfigure for Erroneous or Confusion. Reconfiguration includes step-by-step decomposition, analogy, simplification, and counterexample; failure consists in continuing “as before despite confusion” (Liu et al., 8 Aug 2025).

In the scientific-discovery framework, the middle layer is elevated to the central cognitive act. Model Formation via Qualitative Reasoning is defined as recognizing structural inadequacy, identifying a missing conceptual object, and reformulating the problem within a broader representational space. The paper’s three case studies—Chern’s intrinsic proof of Gauss–Bonnet, Nesterov Accelerated Gradient via Lyapunov functions, and the autonomous disproof of the Erdős unit distance conjecture—are presented as instances where search and execution were insufficient until a new representational object was introduced (Liao, 11 Jun 2026).

A common misconception is that three-layer frameworks are merely pipelines. The literature is less uniform. Some frameworks are strictly sequential, such as MAS-Roles; some are collaborative and dialogic, such as MAS-CMD; some are control loops with regeneration, such as EduStory; and some place the decisive transformation in conceptual reframing rather than in execution.

4. Layer 3: evaluation, human decision, materialization, and feedback closure

The third layer varies most sharply across frameworks, but in each case it closes the system by enforcing correctness, selecting outputs, materializing instruction, or eliciting reflective response. EduStory’s third layer is Learning-Oriented Evaluation Metrics, which quantify knowledge fidelity, pedagogical alignment, and constraint satisfaction. The framework uses Knowledge Drift Rate (KDR),

$S_t = \bigl(E_t,\; R_t,\; C\bigr),$6

Pedagogical Alignment Score (PAS),

$S_t = \bigl(E_t,\; R_t,\; C\bigr),$7

and a Constraint Satisfaction Rate defined from shot-level violations. The same layer includes a Constraint Verifier

$S_t = \bigl(E_t,\; R_t,\; C\bigr),$8

which triggers violation-aware re-generation up to $S_t = \bigl(E_t,\; R_t,\; C\bigr),$9 retries, with fallback to a simplified scene with strict symbol overlays if all retries fail (Wu et al., 10 May 2026).

In ARCHED, the third layer is explicitly human: educators retain control over the workflow, curate AI-generated options, direct revisions, and approve final objectives and downstream assessments. Decision points include defining parameters, reviewing and selecting from LOGS outputs, examining OAE reports, confirming or modifying Bloom-level classifications, iterating, and finalizing objectives. The paper presents this as a response to “black-box” automation, emphasizing downloadable reports and continuous educator input as mechanisms of traceability and accountability (Li et al., 11 Mar 2025).

KLI’s third layer, at the conceptual level, is Instructional Principles: the teaching methods that catalyze the chosen learning processes. In system terms, these principles are selected by the Instructional Principle Agent and then realized by downstream design and feedback agents. In MAS-CMD, this layer is intertwined with persona-specific strategies and a final Decision Agent that selects the most appropriate activity after reviewing the discussion transcript and revised drafts (Wang et al., 20 Aug 2025).

GuideEval’s third layer is Elicitation, which concerns calibrated questioning rather than final selection. It distinguishes E-Strategic, where higher scores are better under Accurate or Comprehension, from E-Heuristic, where lower scores are better under Erroneous or Confusion in order to avoid overload. The layer is thus not merely evaluative; it is metacognitive and state-sensitive (Liu et al., 8 Aug 2025).

IDont materializes the third layer as ContentOntology, which models ContentFragment, ContentObject, and pedagogical ContentType classes such as Facts, Cases, Rules, Models, Theories, together with LearningObject subclasses bound to play, act, scene, and instruction. Content is therefore not an afterthought but a formalized instructional layer aligned to goals and process (Chimalakonda et al., 2018).

A game-based reformulation adds an important dynamic qualification. When the four-layer game framework is consolidated into a three-layer micro–macro–meta model, each layer is said to contain three feedback loops: an exploratory learning loop, an intrinsic motivation loop, and an identity loop. These are formalized as updates for skill/knowledge, motivation, and identity alignment:

$E_t$0

$E_t$1

$E_t$2

This suggests that the third layer in many frameworks is best understood as a closure mechanism: it converts design into assessed action, reflective response, recognized contribution, or executable artifact (Beatty, 2014).

5. Formalization, benchmarks, and empirical evidence

The empirical literature on three-layer frameworks is heterogeneous, ranging from benchmark-driven evaluation to case-study-based deployment. EduStory introduces EduVideoBench, which contains 1,800+ multi-shot STEM instructional clips (30–90 seconds; ≥3 shots) across Physics, Mathematics, Chemistry, Engineering/CS, Biology, Earth/Astronomy, and other topics, with ~100 Manim-rendered clips providing programmatic ground truth for formula correctness and symbolic consistency. On 5,000 held-out shots (Task I) with CogVideoX-2B as base and H100 GPUs, the reported results are: baseline single long prompt with KDR 0.41, PAS 0.52, CLIP-S 0.28; + Instruction Planner (B1) with KDR 0.33, PAS 0.64, CLIP-S 0.29; + Pedagogical State Model (B2) with KDR 0.21, PAS 0.71, CLIP-S 0.28; and EduStory (Full, with verifier) with KDR 0.14, PAS 0.79, CLIP-S 0.27. The paper states that the largest KDR drop occurs when adding state modeling (B1 → B2: −0.12), and that PAS improves by +27 points from baseline to full EduStory (Wu et al., 10 May 2026).

In instructional activity generation, the KLI study evaluates three systems—SAS, MAS-Roles, and MAS-CMD—with 20 practicing teachers and a complementary LLM-as-a-judge system using the Quality Matters (QM) K-12 standards. The reported human-rating results show a significant effect on 5.2 C Active Learning: $E_t$3, with Holm-corrected pairwise comparison MAS-CMD > MAS-Roles (p = .050). The total-score omnibus ANOVA is $E_t$4, while the Friedman test is $E_t$5. The same paper reports substantial computational trade-offs: SAS: 25 ± 16 s; ~3,376 ± 1,601 tokens; 1.00 requests; MAS-Roles: 77 ± 45 s; ~22,497 ± 7,157 tokens; 5.42 ± 1.13 requests; MAS-CMD: 272 ± 168 s; ~71,638 ± 25,222 tokens; 13 requests. Teacher feedback nevertheless strongly preferred MAS-CMD for creativity, contextual relevance, and classroom readiness (Wang et al., 20 Aug 2025).

ARCHED reports two principal empirical evaluations. For analytical capability, it uses 120 existing computer science learning objectives with expert Bloom classification as ground truth and reports weighted Cohen’s Kappa $E_t$6 (95% CI: [0.771, 0.891]), with strongest performance at taxonomy extremes. For generation quality, it compares 30 ARCHED-generated and 30 human-created objectives in computer science, mathematics, and physics using criteria including structural completeness, taxonomic alignment, measurability, content clarity, and technical accuracy, and finds no significant differences (p > 0.05) across all criteria. Reported score examples include structural completeness 4.1±0.4, taxonomic alignment 4.0±0.5, and measurability 3.9±0.5 (Li et al., 11 Mar 2025).

GuideEval contributes a benchmark for tutoring behavior. Its final test set contains 5,177 samples: 1,190 Accurate, 1,181 Erroneous, 1,403 Comprehension, 1,403 Confusion. Automated evaluation uses GPT-4o mini, with reported LLM–human agreement rates and average deviations such as P-Affirm 96.5 / 0.02, P-Redirect 90.0 / 0.05, O-Advance 96.5 / 0.04, and O-Reconfigure 94.5 / 0.06. The paper also reports behavior-guided finetuning of Qwen3-8B with LoRA rank = 8, alpha = 32, max input length = 4096, bfloat16, batch size = 2 per device, gradient accumulation = 8, learning rate = 1e−4, warmup ratio = 0.05, and a single NVIDIA H800 GPU. Under accurate/erroneous conditions, P-Redirect improves from 0.5919 → 0.8890, O-Reconfigure from 0.6681 → 0.8593, O-Advance from 0.9176 → 0.9319, E-Strategic from 2.0605 → 2.1134, and E-Heuristic from 1.9213 → 1.7703 in the sft-think setting (Liu et al., 8 Aug 2025).

IDont does not report quantitative performance metrics. Its principal evidence is feasibility at scale: the framework was used to generate 9 eLearning systems, was designed to support ~1000 similar but varied eLearning systems, and was transferred to the National Literacy Mission of Government of India (Chimalakonda et al., 2018).

6. Limitations, misconceptions, and future directions

Several recurring limitations delimit current three-layer frameworks. EduStory notes restricted domain coverage for subdomains and advanced topics, dependence on script quality, and scalability challenges for videos of many minutes; future directions include richer pedagogical ontologies, robust parsing of imperfect lesson inputs, hierarchical memory, formal verification such as SMT solvers, cross-modal alignment models, and auto-formalization for symbolic verification (Wu et al., 10 May 2026). The KLI study reports small and sometimes statistically insignificant quantitative differences, ceiling effects on some rubric dimensions, very low inter-rater reliability among teachers, reliance on the Gemini family with default decoding settings, and a pronounced compute/latency cost for MAS-CMD (Wang et al., 20 Aug 2025). ARCHED identifies limited validation beyond computer science for higher-order objective generation, adoption barriers in resource-constrained settings, pending LMS integration, and hardware and hosting costs for open-weight LLMs (Li et al., 11 Mar 2025). GuideEval acknowledges that cognitive states are abstracted to four categories, that its data are primarily Chinese middle-school math, that GPT-4o mini evaluation is strong but imperfect, and that ethical tension remains between overguiding and learner autonomy, as well as between politeness and explicit correction (Liu et al., 8 Aug 2025).

The scientific-discovery framework articulates a different limitation: current AI discussion is dominated by search and execution, while Layer 2 is both the most important and the least developed. Its instructional recommendations therefore emphasize analogical transfer drills, cross-disciplinary scanning, concept mapping, abductive reasoning exercises, and temporal awareness of “dormancy periods” in neighboring fields (Liao, 11 Jun 2026). A plausible implication is that future three-layer frameworks may increasingly treat the middle layer not as a routing stage, but as the main locus of pedagogical or scientific intelligence.

A second misconception is that layering necessarily reduces human agency. ARCHED explicitly rejects that assumption by making educators the primary decision-makers, and the KLI literature shows that collaborative multi-agent design can be structured as discussion and decision rather than end-to-end automation (Li et al., 11 Mar 2025, Wang et al., 20 Aug 2025). A third misconception is that formal structure guarantees instructional adequacy. IDont stresses that multiple “correct” ontologies can exist for any domain and that its ontologies are placeholders and exemplars rather than exhaustive specifications (Chimalakonda et al., 2018). A fourth is that visual fluency or surface-level coherence is sufficient in instructional media; EduStory is built precisely on the claim that long-horizon video generation must be not only visually coherent but also pedagogically correct (Wu et al., 10 May 2026).

In aggregate, the literature presents the Three-Layer Instructional Framework as a recurrent method for separating instructional concerns without severing their dependencies. The three layers may be knowledge, learning, and instruction; generation, evaluation, and human decision; state, control, and verification; perception, orchestration, and elicitation; goals, process, and materials; or search, model formation, and execution. What remains stable is the insistence that instructional systems require an explicit representational layer, a mediating adaptation layer, and a closure layer through which correctness, agency, or executable consequence is enforced.