IAAIUniversal YOCO
Universal YOCO for Efficient Depth Scaling
Combines the YOCO decoder-decoder design with recursion: a Universal Self-Decoder iterates with shared parameters, but only over shallow efficient-attention layers, while YOCO's constant global KV cache and linear prefill stay intact.
Inside the method
[ Self-decoder (efficient attention) ] × r → cross-decoder over one global KV cache
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Core recurrence
- Depth control
- Partial recursion in the self-decoder
- KV / state strategy
- Constant global KV cache (YOCO)
Reading note
Remains competitive on general and long-context benchmarks with better token utility and scaling than either YOCO or recursion alone.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{yocou2026,
title = {Universal YOCO for Efficient Depth Scaling},
author = {Yutao Sun and Li Dong and Tianzhu Ye and Shaohan Huang and Jianyong Wang and Furu Wei},
year = {2026},
eprint = {2604.01220},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2604.01220}
}