IAAIDynamic Computation in Looped LMs
Enabling Dynamic Computation in Looped LMs
Points out that Ouro's early exit does not save compute in practice because every depth needs its own KV level. A training-free best-available KV caching strategy lets exited tokens reuse the deepest cache they have, and a fix to the early-exit prior makes tokens exit at genuinely different depths.
Inside the method
Token exits at depth d → later loops read its best-available KV
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Adaptive depth
- Depth control
- Per-token early exit with heterogeneous depths
- KV / state strategy
- Best-available (deepest computed) KV per token
Reading note
Up to 30% lower FLOPs and KV memory while retaining full-depth performance; KV-aware training improves further. Validated on Ouro and smaller looped LMs trained from scratch.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{bestavailablekv2026,
title = {Enabling Dynamic Computation in Looped LMs},
author = {Aayush Mishra and Arnau Padrés Masdemont and Victor Conchello Vendrell and Jordi Ros Giralt and Arash Behboodi and Fabio Valerio Massoli},
year = {2026},
eprint = {2610.09013},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.09013}
}