Jointly predicting courses and grades reduces grade-prediction MAE by 46.4% versus an identical Transformer that predicts grades only¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? A Transformer trained to predict both next-semester course sets and grades achieved MAE 0.1339 versus 0.2496 for a grades-only Transformer, a 46.4% reduction, on a held-out 10% test set. → Paul Savala 2026

Evidence¶

Paul Savala 2026¶

Paul Savala. (2026). Jointly Predicting Courses and Grades Using a Transformer-Based Model. arXiv. https://arxiv.org/abs/2608.13409 doi:10.48550/arxiv.2608.13409

q2 · i? · design · r2

Single-institution experiment training models on 5326 students across 18 semesters, testing on a held-out 10% of students. The results section reports that TRACE achieved a "mean absolute error of 0.1339, a 46.4% reduction" versus the grades-only variant, with MSE 3.5 times lower (0.0392 vs 0.1400).

"The TRACE model achieved a mean absolute error of 0.1339, a 46.4% reduction compared to the ‘OnlyGradesTransformer‘’s MAE of 0.2496, and a mean squared error 3.5 times lower."

Discussion¶

Related Claims¶

• Course prediction achieves modest classification metrics (F1 0.4432) yet the auxiliary task regularizes representations and improves grade prediction — a broader claim this one bears on
• TRACE outperforms LSTM, GNN and XGBoost baselines on grade prediction — related
• KL-divergence and binary cross-entropy losses for course prediction yield grade prediction within 1-2% of each other — related
• Removing student major from the input marginally improves TRACE performance, suggesting major is recoverable from course history — related
• Replacing the transformer with a single-layer 2048-unit LSTM drops the average score by 5.6, with the LSTM winning only on MRPC — related