Fast and Simplex: 2-Simplicial Attention in Triton
Abstract
Recent work has shown that training loss scales as a power law with both model size and the number of tokens, and that achieving compute-optimal models requires scaling model size and token count together. However, these scaling laws assume an infinite supply of data and apply primarily in compute-bound settings. As modern large language models increasingly rely on massive internet-scale datasets, the assumption that they are compute-bound is becoming less valid. This shift highlights the need for architectures that prioritize token efficiency. In this work, we investigate the use of the 2-simplicial Transformer, an architecture that generalizes standard dot-product attention to trilinear functions through an efficient Triton kernel implementation. We demonstrate that the 2-simplicial Transformer achieves better token efficiency than standard Transformers: for a fixed token budget, similarly sized models outperform their dot-product counterparts on tasks involving mathematics, coding, reasoning, and logic. We quantify these gains by demonstrating that -simplicial attention changes the exponent in the scaling laws for knowledge and reasoning tasks compared to dot product attention.
Cite
@article{arxiv.2507.02754,
title = {Fast and Simplex: 2-Simplicial Attention in Triton},
author = {Aurko Roy and Timothy Chou and Sai Surya Duvvuri and Sijia Chen and Jiecao Yu and Xiaodong Wang and Manzil Zaheer and Rohan Anil},
journal= {arXiv preprint arXiv:2507.02754},
year = {2025}
}
Comments
10 pages, with appendix 25 pages