Use of AI in Mathematical Education

Research-Oriented Table of Contents

Part I — What Exactly Is Being Represented?

1. Mathematical knowledge as a carrier

1.1 Standard representation: curriculum as ordered topics; hidden assumptions: monotone progression, stable prerequisites, one dominant decomposition.
1.2 Knowledge graphs: nodes = concepts, edges = prerequisite/implication/dependence.
1.3 Learning spaces: admissible knowledge states $K\subseteq 2^Q$, allowing multiple legal paths through the same domain.
1.4 Hypergraph representations: prerequisites may be conjunctive, disjunctive, compensatory, or context-dependent.
1.5 Concept lattices and Galois structures: concepts represented through shared attribute closure rather than curricular sequence.
1.6 Equivalence boundary: sequence $\simeq$ DAG only when prerequisite relation is effectively total or near-total; DAG $\not\simeq$ learning space when alternative acquisition paths matter.
1.7 Degenerate cases: one-concept domains, perfectly hierarchical domains, completely independent skills.
1.8 Extremal cases: massive prerequisite networks; highly entangled concepts; cyclic dependence.
1.9 Local/global fracture: local prerequisite validity need not compose into globally coherent curriculum structure.
1.10 Easy-after-representation-change: course planning becomes reachability on a state space rather than heuristic sequencing.
1.11 Hidden structure: alternative learning paths, redundancy, bottleneck concepts, articulation points, prerequisite sparsity.
1.12 Neighboring fields: graph theory, formal concept analysis, planning, knowledge-space theory, causal DAGs.
1.13 Reopened formulations: programmed instruction, mastery learning, knowledge-space theory, adaptive hypermedia.
1.14 Open problem: whether AI-generated curricular maps discover genuinely new admissible paths or merely reconstruct canonical textbook order.

Unusual branch — curriculum as reachable-state geometry.
what changes: topics become states and transitions. what becomes visible: alternative legal trajectories and bottlenecks. what becomes harder: state-space identification. motivation: knowledge-space theory and graph reachability. test: compare linear-sequence tutoring against state-space navigation under equal instructional budget.

---

Part II — Learner State Is Not a Score

2. Representations of mathematical competence

2.1 Standard representation: scalar score, percent correct, latent ability $\theta$.
2.2 Hidden assumptions: unidimensionality, stationarity, local independence, score sufficiency.
2.3 Bayesian knowledge tracing: $P(K_t\mid y_{1:t})$.
2.4 Cognitive-diagnostic models: mastery vector $z\in\{0,1\}^d$.
2.5 Dynamic state-space models: mastery, forgetting, interference, switching strategies.
2.6 Misconception graphs: nodes encode erroneous generative rules rather than missing skills.
2.7 Causal learner models: intervention $a_t$ changes latent state $s_t\to s_{t+1}$.
2.8 Equivalence boundary: scalar ability approximates mastery vector only when one latent direction dominates.
2.9 Failure boundary: compensatory skills, conjunctive tasks, guessing, forgetting, context-dependent strategy choice.
2.10 Singular cases: perfect scores with brittle transfer; low scores caused by representation rather than knowledge.
2.11 High-dimensional regime: many weakly observed skills, severe identifiability problems.
2.12 Easy-after-change: “Which student is weak?” becomes “Which latent configuration produces this trace?”
2.13 Hidden invariant: identical correctness can arise from non-equivalent internal states.
2.14 Neighboring fields: psychometrics, hidden Markov models, system identification, causal inference, partially observable control.
2.15 Open problem: distinguish learner models with equivalent predictive accuracy but different optimal interventions.

Unusual branch — intervention-equivalence classes.
what changes: learner models are judged by induced pedagogical policy, not prediction loss. what becomes visible: models that predict identically but prescribe differently. what becomes harder: counterfactual validation. motivation: causal decision theory. test: hold predictive accuracy fixed; compare transfer under interventions chosen by competing models.

---

Part III — Mathematical Objects Have Multiple Educational Carriers

3. Symbolic, graphical, executable, and formal representations

3.1 Standard representation: prose + symbolic notation.
3.2 Hidden assumption: notation is a neutral wrapper around invariant mathematical content.
3.3 Graphical/dynamic representations.
3.4 Tabular/numerical representations.
3.5 Diagrammatic and topological representations.
3.6 Programmatic/executable representations.
3.7 Formal proof objects and proof states.
3.8 Manipulative, embodied, and geometric representations.
3.9 Equivalence condition: representations are equivalent only relative to preserved distinctions $I$: $r_1\simeq_I r_2$.
3.10 Failure boundary: information lost by discretization, projection, numerical approximation, symbol compression, dimensional reduction.
3.11 Boundary cases: discrete/continuous, exact/approximate, local/global, finite/infinite, low/high dimension.
3.12 Easy-after-change: recurrence $\to$ state machine; algebraic identity $\to$ area decomposition; transformation $\to$ dynamic geometry; combinatorial object $\to$ graph.
3.13 Hidden structures: symmetry, conservation, sparsity, duality, invariant subspace, monotonicity.
3.14 Neighboring fields: visualization, HCI, programming languages, formal methods, scientific computing.
3.15 Open problem: infer $r^*=\arg\max_r \text{learning gain}(r\mid learner,object,state)$.

Unusual branch — representation policy instead of explanation policy.
what changes: AI chooses the carrier before generating explanation. what becomes visible: invariants suppressed by the current representation. what becomes harder: measuring faithful transport between carriers. motivation: multiple-representation learning and semiotic-register theory. test: compare fixed-representation tutoring with adaptive carrier switching.

---

Part IV — Error as Surface Failure versus Generative Mechanism

4. Representing mathematical mistakes

4.1 Standard representation: correct/incorrect.
4.2 Step-level error labels.
4.3 Misconception taxonomies.
4.4 Minimal counterexample representation.
4.5 Error dependency graph.
4.6 Minimal causal failure set $\varkappa$: smallest assumption/transition whose removal changes the failure.
4.7 Equivalence boundary: outcome correctness reflects reasoning only when solution routes are essentially unique.
4.8 Failure cases: lucky guesses, compensating errors, multiple valid strategies, shared wrong answers generated by distinct rules.
4.9 Local/global fracture: local arithmetic mistakes versus global representational misunderstanding.
4.10 Deterministic/stochastic errors: stable misconception versus random execution noise.
4.11 Easy-after-change: repetitive “wrong answer” patterns become one reusable causal diagnosis.
4.12 Neighboring fields: program debugging, fault localization, causal diagnosis, cognitive architectures.
4.13 Open problem: whether LLMs can infer stable generative misconceptions rather than merely classify surface errors.

Unusual branch — counterkernel tutoring.
what changes: feedback targets minimal failure owner. what becomes visible: causal error structure. what becomes harder: nonuniqueness of minimal explanations. motivation: debugging and minimal-unsatisfiable-subset methods. test: measure recurrence under isomorphic-but-reworded problems after ordinary versus counterkernel feedback.

---

Part V — Proof as Text versus Executable State

5. AI and proof education

5.1 Standard representation: final natural-language proof.
5.2 Proof tree.
5.3 Proof DAG with shared lemmas.
5.4 Tactic-state trajectory.
5.5 Lean/Coq proof term.
5.6 Counterexample-guided proof search.
5.7 Informal concept map + formal kernel.
5.8 Equivalence boundary: informal and formal proof agree only after faithful transport of definitions, hidden assumptions, and inference rules.
5.9 Failure boundary: formally valid but pedagogically opaque proof; intuitively clear but formally incomplete proof.
5.10 Low-dimensional case: one-step derivation.
5.11 Extremal case: massive formal proof with little global human comprehension.
5.12 Easy-after-change: justification checking becomes executable rather than rhetorical.
5.13 Hidden structure: reusable lemma graph, dependency bottlenecks, unnecessary assumptions.
5.14 Neighboring fields: theorem proving, program synthesis, proof-carrying code, type theory.
5.15 Open problem: whether small LLM + strong proof kernel can outperform frontier LLM tutoring.

Unusual branch — proof-state pedagogy.
what changes: tutor observes goals/context/tactics instead of prose alone. what becomes visible: exact unresolved obligations. what becomes harder: converting kernel errors into meaningful pedagogy. motivation: interactive theorem proving. test: compare proof-state-conditioned hints against transcript-only hints on delayed unaided proofs.

---

Part VI — AI Tutor Is the Wrong Primitive

6. From conversational agent to pedagogical control system

6.1 Standard representation: learner $\leftrightarrow$ chatbot.
6.2 Hidden assumption: language generation policy is equivalent to pedagogical policy.
6.3 Diagnosis → pedagogical intent → representation selection → generation.
6.4 Model-tracing tutor + LLM surface realization.
6.5 Constraint-based tutor + LLM repair.
6.6 Multi-agent generator/critic/verifier architecture.
6.7 Planner over learner state $s_t$, action $a_t$, outcome $y_t$.
6.8 Equivalence boundary: monolithic LLM approximates controlled tutoring only when nearly all plausible responses are pedagogically admissible.
6.9 Failure boundary: answer leakage, premature hints, over-explanation, incorrect diagnosis, excessive dependency.
6.10 Local/global distinction: next hint versus curriculum trajectory.
6.11 Neighboring fields: control theory, POMDPs, adaptive systems, intelligent tutoring systems.
6.12 Historically abandoned branch: rule-based ITS architectures may become useful again with LLM language layers.
6.13 Open problem: separate language competence from pedagogical policy quality.

Unusual branch — LLM as actuator, not controller.
what changes: decision policy is external to generation. what becomes visible: pedagogical failure can be localized separately from language failure. what becomes harder: controller design. motivation: intelligent tutoring systems and control architectures. test: same LLM, different controller; compare learning outcomes.

---

Part VII — Multi-AI Exploration as Search Geometry

7. Ensembles beyond voting

7.1 Standard representation: several models answer; majority agreement increases confidence.
7.2 Hidden assumption: errors are sufficiently independent.
7.3 Structural-role decomposition: generator, counterexample finder, formalizer, critic, verifier, curriculum mapper.
7.4 Independent topic maps.
7.5 Nonaliased route detection: $R_i\neq R_j$ only if load-bearing structure differs.
7.6 Agreement versus diversity: consensus measures density; structural disagreement exposes frontier.
7.7 Equivalence boundary: ensemble voting works when failures are independent and objectives aligned.
7.8 Failure boundary: correlated training distributions, shared misconceptions, identical search basins.
7.9 Easy-after-change: next-step generation becomes frontier expansion rather than averaged recommendation.
7.10 Neighboring fields: ensemble methods, adversarial search, scientific discovery systems, multi-agent planning.
7.11 Open problem: maximize $\Delta Reach$, not model count.

Unusual branch — nonconsensus educational ensemble.
what changes: agents are rewarded for structurally distinct diagnoses. what becomes visible: competing learner models. what becomes harder: arbitration. motivation: adversarial validation. test: majority-vote versus structurally diversified ensemble under equal compute.

---

Part VIII — Assessment as Invariance, Not Immediate Performance

8. What counts as learning under AI assistance?

8.1 Standard representation: immediate pre/post score.
8.2 Delayed retention.
8.3 Cross-representation transfer.
8.4 AI-withdrawal assessment.
8.5 Adversarially perturbed problems.
8.6 Novel-context transfer.
8.7 Strategy reconstruction.
8.8 Human–AI joint performance versus unaided performance.
8.9 Equivalence boundary: score gain $\simeq$ learning only when task distribution and scaffolding match target competence.
8.10 Failure boundary: answer production improves while independent strategy selection deteriorates.
8.11 Extremal case: near-perfect assisted performance with zero unaided transfer.
8.12 Hidden invariant: competence should survive transformation $g\in G$ of surface representation.
8.13 Proposed criterion: $C(x)=\Pr[\text{success under unseen }g(x),\ \text{AI removed}]$.
8.14 Neighboring fields: robustness, invariance learning, causal inference, domain adaptation.
8.15 Open problem: identify transformations that preserve mathematical competence but destroy superficial memorization.

Unusual branch — competence as an invariance class.
what changes: ability becomes stability under transformations. what becomes visible: scaffold dependence. what becomes harder: constructing valid transformations. motivation: invariance and transfer theory. test: train in one representation, test in unseen representations with assistance removed.

---

Part IX — AI-Generated Mathematical Worlds

9. Synthetic examples, counterexamples, and problem spaces

9.1 Standard representation: fixed textbook problem bank.
9.2 Procedural generation.
9.3 Symbolic problem generators.
9.4 LLM-generated semantic variants.
9.5 Counterexample-generating systems.
9.6 Parameterized families rather than isolated problems.
9.7 Equivalence boundary: two problems are educationally equivalent only relative to preserved conceptual structure.
9.8 Failure boundary: surface diversity with identical solution template.
9.9 Singular/extremal-case generation.
9.10 High-dimensional and stochastic problem environments.
9.11 Easy-after-change: one exercise becomes a structured neighborhood of perturbations.
9.12 Hidden structure: sensitivity of solution method to assumptions.
9.13 Neighboring fields: property-based testing, fuzzing, benchmark generation, automated theorem proving.
9.14 Open problem: generate problems maximizing conceptual discrimination rather than difficulty.

Unusual branch — mathematical fuzzing for education.
what changes: problem generation searches boundaries where student models disagree. what becomes visible: brittle understanding. what becomes harder: preserving mathematical validity. motivation: software fuzzing/property testing. test: compare random variants with boundary-targeted variants for misconception detection.

---

Part X — Historically Marginal Systems Reopened by Modern AI

10. Old architectures with new computational economics

10.1 Programmed instruction.
10.2 Mastery learning.
10.3 LOGO and constructionist microworlds.
10.4 Computer algebra as pedagogy.
10.5 Knowledge-space theory.
10.6 Model-tracing tutors.
10.7 Constraint-based tutoring.
10.8 Formal theorem proving for students.
10.9 Cognitive architectures.
10.10 Automated curriculum generation.
10.11 Why these approaches previously failed: authoring cost, brittleness, sparse interfaces, computational cost.
10.12 Why LLMs alter the economics: interface generation, automatic formalization, explanation generation, synthetic examples, translation between carriers.
10.13 Open problem: determine which abandoned approaches failed structurally and which failed only because authoring/interface costs were too high.

Unusual branch — resurrected symbolic pedagogy.
what changes: symbolic engines regain natural-language interfaces. what becomes visible: explicit pedagogical state hidden by pure LLM chat. what becomes harder: integrating symbolic rigidity with open-ended dialogue. motivation: classical ITS. test: rebuild historically successful but expensive architectures with LLM-generated interface/authoring and compare equal-budget outcomes.

---

Part XI — Open Problems That May Be Representation-Dependent

11.1 Is “mastery” fundamentally scalar, vector-valued, graph-valued, or policy-relative?
11.2 Are misconceptions better represented as missing knowledge, alternative rules, or causal generators?
11.3 Is the right tutoring action an explanation, representation change, counterexample, proof obligation, or deliberate silence?
11.4 Can formal proof states outperform natural-language traces as learner-state sensors?
11.5 Can small models plus strong external discriminators outperform very large conversational models?
11.6 Can AI discover noncanonical curricular paths inaccessible to human-authored prerequisite graphs?
11.7 Can mathematical understanding be operationalized as invariance under representation change?
11.8 Which AI benefits persist after removal of AI?
11.9 When does computational representation create understanding versus merely relocate cognitive work?
11.10 Can multi-AI disagreement reveal latent ambiguity in the learner rather than model noise?
11.11 Can adaptive systems discover new pedagogical primitives rather than optimize existing ones?
11.12 Does the fundamental unit of mathematics education become student × representation × tool × task rather than “student ability”?

Representation Fractures

1. Curriculum sequence → learning-space topology. The reachable object changes from “next topic” to “set of admissible successor knowledge states.” Alternative routes, bottlenecks, compensatory paths, and local/global prerequisite failures become explicit.

2. Score → causal learner state. Two students with identical scores cease to be equivalent. The research space changes from prediction to intervention: $P(y_{t+1})$ is replaced by $P(s_{t+1}\mid do(a_t))$.

3. Static notation → executable mathematical carrier. A formula becomes something that can be run, perturbed, visualized, formally checked, and adversarially tested. Entire classes of experimentation become reachable.

4. Natural-language reasoning → proof-state/dependency structure. Correctness, unresolved obligation, hidden assumption, lemma reuse, and failure localization become machine-observable rather than rhetorically inferred.

5. Chatbot → closed-loop pedagogical control system. The primitive changes from prompt → response to state estimate → representation choice → intervention → trace → discriminator → state update. This is the largest fracture because it makes the LLM only one component inside the educational system rather than the educational system itself.