TAPS Scheduler Optimizes Scaling in Recursive Reasoning Transformers
September 28, 2026
The Trajectory Adaptive Progress-Fluctuation Scheduler (TAPS) dynamically adjusts the scale of recurrent updates in looped transformers. By decomposing terminal loss into persistent-progress and centered-fluctuation components, TAPS prevents overly conservative or aggressive updates during test-time computation scaling.
HOW THIS AFFECTS YOU
●
builderThis provides a more efficient way to scale test-time computation for reasoning-heavy tasks.
●
researcherYou can improve the stability of recurrent reasoning models by adapting update scales to the latent trajectory.