Work

YS Learner: A Deterministic, Exam-Aware, Multidimensional Adaptive Learning Architecture

Year: 2026
Type: article
Abstract: High-stakes competitive exam preparation is a tightly constrained resource-allocation problem. Every day, a candidate balances a finite study budget and an immovable exam date against uneven syllabus mark weights. At the same time, the learner’s internal cognitive state is constantly shifting—driven by new concept acquisition, exponential memory decay, and recurring diagnostic mistakes. Standard question banks reduce this multidimensional problem to backward-looking accuracy percentages and unsorted lists of weak topics. They fail to answer the learner’s most urgent daily question: what exact study action should receive today’s limited cognitive capacity, in what format, and why? We built YS Learner to solve this decision bottleneck directly on the learner’s device. Rather than relying on opaque cloud neural recommenders or fragile scalar weighted formulas, we engineered a deterministic, exam-aware adaptive engine in pure, Android-decoupled JVM/Kotlin (com.psuece.practice.*). We enforce strict boundaries across six architectural layers so that upstream behavioral telemetry cannot silently corrupt downstream study schedules. Specifically, our architecture introduces: (1) confidence-qualified evidence weighting (𝑤𝑖 ∈ [0, 1]); (2) dual-horizon, credibility-adjusted mastery scoring (𝑀𝑉2) anchored by a small-sample linear ramp (𝐾 = 5.0); (3) exponential memory decay tracking (𝑅(Δ𝑡)) governed by a 7.0-day half-life; (4) a noncompensatory 9-tier lexicographic optimizer paired with largest-remainder integer difficulty reconciliation; (5) a 45-category 5S diagnostic error taxonomy that isolates 39 learner breakdown mechanisms from 6 content-defect quarantine guards; and (6) phase-aware lifecycle governance across five preparation tracks (R360, F180, C90, P60, and CLASSROOM_360). We verified the underlying pure-JVM implementation through unit testing, static architectural boundary guards, and multi-phase lifecycle simulations, achieving a 437/442 baseline test pass rate, 84/84 traceability coverage, and 15/15 mutation kills. Prospective, longitudinal validation of examination pass rates remains a separate empirical task outlined in Section 8. Keywords: Adaptive Learning Engines, Deterministic Optimization, Spaced Repetition Systems, Item Response Theory, Error Diagnostics, Edge Intelligence, Offline-First Mobile Architecture.
Author: Borah, Moloy
Language: English
Open Access status: green
OpenAlex ID: W7236594822
FWCI:
Cites:
Cited by:
Related to: 10