IAAITaH2
Improving Test-Time Scaling with Adaptive Looped Transformers
Studies whether looping improves test-time scaling as outputs grow longer and finds fixed-depth looping wastes iterations on tokens that do not benefit. TaH2 post-trains the backbone with an iteration decider supervised by online lookahead labels indicating whether another iteration improves the prediction.
Inside the method
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Adaptive depth
- Depth control
- Token-level adaptive iterations up to a maximum depth
- KV / state strategy
- See paper
Reading note
On AIME it improves the accuracy-compute slope by 53% over the non-looped baseline and exceeds its peak accuracy by about 3.4 points at matched compute. Gains over the baseline grow from +2.8 at depth 2 to +3.9 at depth 8.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{tah22026,
title = {Improving Test-Time Scaling with Adaptive Looped Transformers},
author = {Yichen You and Tianyu Fu and Aosong Feng and Xingtai Lv and Xuefei Ning and Ning Ding and Yu Wang},
year = {2026},
eprint = {2609.35748},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.35748}
}