FPRT Method Reduces Distribution Mismatch in Reasoning Distillation
November 9, 2026
The Failure Profiling and Reachable Trajectory Selection (FPRT) method aims to stop small models from overfitting to teacher model output distributions. It selects training data based on a student model's specific failure patterns and its ability to realistically learn those trajectories.
HOW THIS AFFECTS YOU
●
builderYou can improve how small models learn reasoning by training them on data that matches their own actual capabilities.
●
researcherYou can study how student-specific failure profiling helps solve distribution mismatch during distillation.