PRISM: Parallel Residual Iterative Sequence Model
Jie Jiang, Ke Cheng, Xin Xu, Mengyang Pang, Tianhao Lu, Jiaheng Li, Yue Liu, Yuan Wang, Jun Zhang, Huan Yu, Zhouchen Lin · 2026
Read PRISM: Parallel Residual Iterative Sequence Model by Jie Jiang, Ke Cheng, Xin Xu, Mengyang Pang, Tianhao Lu, Jiaheng Li, Yue Liu, Yuan Wang, Jun Zhang, Huan Yu, Zhouchen Lin (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-10796.txt.
Description
Jie Jiang and ten co-authors match expressivity with parallelism in sequence modeling in a 2026 paper. Transformers are expressive but costly; linear sequence models are fast but bounded by shallow, single-step updates. Iterative methods like test-time training add power yet break hardware parallelism through token-level and step-level serial dependencies.
Their model, PRISM, approximates the gate-residual-direction iteration pattern of test-time training in a parallelizable form. A write-forget decoupling plan isolates nonlinearity within the injection rule, while a two-stage proxy system sidesteps stated solvers' serial dependency. A short convolution anchors the initial residual, letting the model iterate in parallel. The design aims at sequence models that are both strong and hardware-friendly.
Catalog record
| CREATOR | Jie Jiang, Ke Cheng, Xin Xu, Mengyang Pang, Tianhao Lu, Jiaheng Li, Yue Liu, Yuan Wang, Jun Zhang, Huan Yu, Zhouchen Lin |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 75 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-10796