NN-Former: Rethinking Graph Structure in Neural Architecture Representation
Ruihan Xu, Haokui Zhang, Yaowei Wang, Wei Zeng, Shiliang Zhang · 2025
Read NN-Former: Rethinking Graph Structure in Neural Architecture Representation by Ruihan Xu, Haokui Zhang, Yaowei Wang, Wei Zeng, Shiliang Zhang (2025) 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-2507-00880.txt.
Description
Ruihan Xu and co-authors present NN-Former in a 2025 paper on predicting how a neural network will perform before it is trained. Designers rely on such predictors to judge accuracy and speed cheaply. Graph-based predictors miss rich features, while transformer-style ones falter as networks grow deeper, so the team rethought how a network's structure is encoded.
Their key insight is that sibling nodes, layers fed by the same parent, matter more than past work assumed. NN-Former blends a graph network with a transformer and adds a mixer that pays attention to siblings. A second new part passes messages both ways along the graph. Across tests, the predictor led rivals on both accuracy and speed estimates. The authors share their code openly.
Catalog record
| CREATOR | Ruihan Xu, Haokui Zhang, Yaowei Wang, Wei Zeng, Shiliang Zhang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-00880