VecFormer: Towards Efficient and Generalizable Graph Transformer with Graph Token Attention Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Arab Emirates DOI:

Jingbo Zhou † \dagger, Jun Xia † \dagger, Siyuan Li † \dagger, Yunfan Liu, Wenjun Wang, Yufei Huang, Changxi Chi, Mutian Hong, Zhuoli Ouyang, Shu Wang, Zhongqi Wang, Xingyu Wu, Chang Yu † \dagger, Stan Z. Li · 2026

Read VecFormer: Towards Efficient and Generalizable Graph Transformer with Graph Token Attention Conference: Proceedings of the ACM Web Conference 2026; April 13–17, 2026; Dubai, United Arab Emirates Proceedings of the ACM Web Conference 2026 (WWW ’26), April 13–17, 2026, Dubai, United Arab Emirates DOI: by Jingbo Zhou † \dagger, Jun Xia † \dagger, Siyuan Li † \dagger, Yunfan Liu, Wenjun Wang, Yufei Huang, Changxi Chi, Mutian Hong, Zhuoli Ouyang, Shu Wang, Zhongqi Wang, Xingyu Wu, Chang Yu † \dagger, Stan Z. Li (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-2602-19622.txt.

Description

Graph Transformers shine at graph representation learning, yet most models blow up in compute as graphs grow, and node-level attention generalizes poorly out of distribution. Jingbo Zhou and thirteen co-authors propose VecFormer, the Vector Quantized Graph Transformer, built for node classification under out-of-distribution shifts. It trains in two stages, first learning rich Graph Codes from two codebooks, then attending at the Graph Token level.

Stage one rebuilds node features and graph structure from two codebooks, aiming to learn rich semantic Graph Codes. Stage two runs attention on Graph Tokens via the transformed cross codebook, cutting compute while lifting generalization. Tests on varied-size data sets show VecFormer beating current Graph Transformers on both speed and scores.

Catalog record

CREATORJingbo Zhou † \dagger, Jun Xia † \dagger, Siyuan Li † \dagger, Yunfan Liu, Wenjun Wang, Yufei Huang, Changxi Chi, Mutian Hong, Zhuoli Ouyang, Shu Wang, Zhongqi Wang, Xingyu Wu, Chang Yu † \dagger, Stan Z. Li
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-19622

Browse the full library · About the Tower of Babel