IAAICHASE
CHASE: Cache-Hole-Adapted Skip Exit for Looped State-Space Language Models
Extends looping to state-space language models with Looped Mamba and looped hybrid Mamba-Transformer blocks, and adapts Ouro-style exit gates. Because early exits leave the recurrent SSM state without deeper updates, a cache-hole adaptation aligns continued training with skipped-state inference.
Inside the method
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Adaptive depth
- Depth control
- Threshold-controlled early exit
- KV / state strategy
- SSM state with skipped deeper updates
Reading note
Looped Mamba beats parameter-matched baselines on controlled reasoning tasks; in pretraining it stays competitive downstream with fewer parameters, though deeper non-looped models keep a perplexity edge at strict iso-FLOPs. Adapted exits keep perplexity near full computation.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{chase2026,
title = {CHASE: Cache-Hole-Adapted Skip Exit for Looped State-Space Language Models},
author = {Zhenxuan Yu and Takeshi Kojima and Yutaka Matsuo and Yusuke Iwasawa},
year = {2026},
eprint = {2607.10110},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2607.10110}
}