Context-Aware Learning Curve Extrapolation with Prior-Data Fitted Networks

Extrapolating learning curves of machine learning algorithms can help estimate future performance at larger training set sizes and inform data collection budgets, yet remains challenging on heterogeneous tabular benchmarks where curve shapes vary across learners and tasks. We revisit real-data training for Learning Curve Prior-Data Fitted Networks and find that direct training often overfits. We first introduce Real-LC-PFN++, a stabilized real-data baseline that combines early stopping and a content-free learnable dummy token (also called register token). Building on this, we propose CTX-LC-PFN, a context-aware extrapolator that conditions its predictive distribution on auxiliary context, specifically, the learner identity and observed training performance. In the largest learning curve benchmark, LCDB 1.1, our methods improve uncertainty calibration and extrapolation accuracy across observation regimes. To improve accessibility and reproducibility, we will release code, trained models, and data processing pipelines upon acceptance.