This study presents a systematic review and empirical analysis of Knowledge Tracing (KT) models from 2022 to 2025. Based on architectural and design principles, recent KT models are categorized into six directions: dynamic memory architectures, graph-structured modeling, attention mechanisms, explainability-oriented approaches, multi-relational reasoning, and automated personalization via representation learning. Multiple representative models were re-implemented in a unified environment with behavioral feature stratification (item difficulty, learner ability, response speed, hint usage) for cross-context testing. Results show attention-based models maintain the highest stability under multi-feature interference, while multi-relational memory models excel in structured, high-pressure contexts. Future work includes expanding evaluation to more diverse datasets, developing architectures with dynamic feature detection and adaptation, and broadening behavioral analysis to additional real-world learning signals.