From Table to Cell: Attention for Better Reasoning with TabAlign
Tung Sum Thomas Kwok, Zeyong Zhang, Xinyu Wang, Chunhe Wang, Xiaofeng Lin, Hanwei Wu, Lei Ding, Guang Cheng, Zhijiang Guo · 2026
Read From Table to Cell: Attention for Better Reasoning with TabAlign by Tung Sum Thomas Kwok, Zeyong Zhang, Xinyu Wang, Chunhe Wang, Xiaofeng Lin, Hanwei Wu, Lei Ding, Guang Cheng, Zhijiang Guo (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-2605-14465.txt.
Introduction
Kwok and colleagues argue multi-step table reasoning fails without a clear cell-grounding contract between planning and execution. Older methods force a left-to-right planner factorisation at odds with permutation invariance, and score states by content alone. A pilot study shows diffusion models produce more order-stable cell attention than autoregressive models. The team presents TABALIGN, pairing a diffusion planner that turns plans into cell masks with TABATTN, a checker scoring steps by mask overlap. Across eight benchmarks, TABALIGN beats the strongest baseline by 15.76 percentage points on accuracy.
Catalog record
| CREATOR | Tung Sum Thomas Kwok, Zeyong Zhang, Xinyu Wang, Chunhe Wang, Xiaofeng Lin, Hanwei Wu, Lei Ding, Guang Cheng, Zhijiang Guo |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 136035 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.14465