AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Hyper-SET: Designing Transformers via Hyperspherical Energy Minimization

Yunzhe Hu, Difan Zou, Dong Xu

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Transformer expressivity and learnability · 4 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use bi-softmax attention instead of sigmoid attention (ablation)Hyper-SET · CIFAR-10top-1 accuracy (%): 85.93 -> 90.11Improved
use bi-softmax attention instead of linear attention (ablation)Hyper-SET · CIFAR-100top-1 accuracy (%): 56.97 -> 63.41Improved
use ReLU feedforward instead of softmax feedforward (ablation)Hyper-SET · CIFAR-10top-1 accuracy (%): 88.20 -> 90.11Improved
use ReLU feedforward instead of gated feedforward (ablation)Hyper-SET · CIFAR-100top-1 accuracy (%): 59.29 -> 63.41Improved
learn step sizes instead of fixing them to 0.5 (ablation)Hyper-SET · CIFAR-10top-1 accuracy (%): 25.81 -> 90.11Improved
learn step sizes instead of fixing them to 0.1 (ablation)Hyper-SET · CIFAR-10top-1 accuracy (%): 81.45 -> 90.11Improved
add depth-wise LoRA rank 4 (ablation)Hyper-SET · ImageNet-100top-1 accuracy (%): 70.16 -> 70.36Improved
add depth-wise LoRA rank 32 (ablation)Hyper-SET · ImageNet-100top-1 accuracy (%): 70.16 -> 72.20Improved
scale up hidden dimension to match Transformer parametersHyper-SET · ImageNet-100top-1 accuracy (%): 67.64 -> 70.16Improved
scale up iterations to 24 and feedforward width to 8dHyper-SET · ImageNet-100 masked image modeling (40% mask)FID: 59.841 -> 45.174Improved
use Hyper-SET instead of Transformer with fewer parametersSudoku · Sudokuaccuracy (%): 49.30 -> 54.70Improved
use Hyper-SET Small instead of Transformer on CIFAR-10Image classification · CIFAR-10top-1 accuracy (%): 89.90 -> 88.89Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy