TAPS scheduler adaptively scales recurrent updates in looped transformers, up to 1.56x speedup

Boyuan Wang · hf · 2026-10-01

Recurrent reasoning models apply updates at a fixed scale — conservative during persistent progress, aggressive during fluctuation. TAPS (Trajectory Adaptive Progress-Fluctuation Scheduler) decomposes loss sensitivity to update scale into progress and fluctuation terms and adapts step size online, with theoretical guarantees. Empirically it improves terminal accuracy on structured reasoning without retraining, and yields up to 1.56x wall-clock speedup at matched accuracy when incorporated into training, across diverse recurrent architectures.

Original post →

More from Research

Research channel →