IAAITAPS
Scheduling Recursive Reasoning in Looped Transformers
Adapts the scale of each recurrent update online. The Trajectory Adaptive Progress-Fluctuation Scheduler decomposes sensitivity of the terminal loss into persistent progress and fluctuation terms and adjusts the step size accordingly, at inference without retraining or also during training.
Inside the method
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Adaptive depth
- Depth control
- Fixed loops with adaptive update scale
- KV / state strategy
- See paper
Reading note
Improves terminal accuracy on structured reasoning tasks across several recurrent architectures; when used in training, reaches baseline accuracy with up to 1.56× wall-clock speedup. Provides sufficient conditions for fewer loops to a target quality.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{taps2026,
title = {Scheduling Recursive Reasoning in Looped Transformers},
author = {Boyuan Wang and Chengyao Yu and Jiaxi Ren and Hongxin Wei and Bingyi Jing and Yuxin Tao},
year = {2026},
eprint = {2609.36653},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.36653}
}