AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

OLion: Approaching the Hadamard Ideal by Intersecting Spectral and L inf Implicit Biases

Zixiao Wang, Yifei Shen, Huishuai Zhang

ICML 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: Low-rank compression and merging

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
apply sign after Newton-Schulz orthogonalization (OLion) instead of Muon orthogonalization aloneLlama-3.1-8B SFT from AdamW-pretrained checkpoint · GSM8K 0-shotaccuracy (%): 57.24 -> 60.04Improved
apply sign after orthogonalization (OLion) vs MuonLlama-3.1-8B SFT · Aqua 0-shotaccuracy (%): 41.34 -> 46.45Improved
OLion vs AdamW for SFT of an AdamW-pretrained modelLlama-3.1-8B SFT · GSM8K 4-shotaccuracy (%): 60.35 -> 60.58No clear effect
OLion vs Muon on MATH 0-shot SFTLlama-3.1-8B SFT · MATH 0-shotaccuracy (%): 19.32 -> 19.80No clear effect
use OLion instead of AdamW, Lion and Muon for GPT-2 pretrainingGPT-2 124M-770M on OpenWebText · validation lossconvergence speedImproved
use OLion for Llama-2-7B pretrainingLlama-2-7B · training and validation losslossImproved
vary learning rate (3e-4 to 5e-3) with OLion (ablation)GPT-2 small · validation loss at step 10,000validation lossImproved
use OLion instead of Muon for SiT-B/2 pretrainingSiT-B/2 on ImageNet-1K · projection and denoising losslossImproved
add sign step after orthogonalizationGPT-2 small weights · spectral and l-inf normsl-inf normImproved

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