IAAIDual-Slot
Dual-Slot: From Serial Loops to Parallel Slots in Language Models
Reorganizes serial looped computation into parallel latent and prediction slots. The prediction slot uses the previous token's deep latent state, while the latent slot prepares the state for the next step, replacing two sequential invocations with one paired call.
Inside the method
Token t: prediction slot (deep latent of t−1) ‖ latent slot (state for t+1)
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Full-stack recurrence
- Depth control
- One paired invocation per token
- KV / state strategy
- See paper
Reading note
On Qwen-style 1B dense models, Pile perplexity drops from 7.25 to 7.00 and the five-shot average rises 2.26 points. Training uses a few Jacobi iterations; decoding keeps 93.1–96.5% of vanilla throughput for the dense model.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{dualslot2026,
title = {Dual-Slot: From Serial Loops to Parallel Slots in Language Models},
author = {Yiqin Hao and Zitong Wang and Boyi Zeng and He Li and Shixiang Song and Feichen Song and Yifan Liu and Ran Tao and Bryan Dai and Zhouhan Lin},
year = {2026},
eprint = {2610.25086},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.25086}
}