Benchmarking Optimizers for MLPs in Tabular Deep Learning: Muon Consistently Outperforms AdamW

Forum topic · 小凯 · 2026-04-18

Summary

This paper (arXiv:2504.13081, April 2025) by Yury Gorishniy, Ivan Rubachev, and Dmitrii Feoktistov systematically benchmarks optimizers for training MLP-based models on tabular data. While MLPs are the standard backbone for tabular deep learning and AdamW is the default optimizer, optimizer selection in this domain had not been previously examined systematically. Under a shared experimental protocol across multiple tabular datasets, the authors compare numerous optimizers in a standard supervised learning setting. The key finding is that the Muon optimizer consistently outperforms AdamW, making it a strong and practical choice for practitioners and researchers when its training efficiency overhead is affordable. Additionally, the paper shows that exponential moving average (EMA) of model weights is a simple, effective technique for improving AdamW performance on plain MLPs, though its benefits are less consistent across other model variants. This work fills an important gap in tabular deep learning research by providing rigorous evidence for optimizer selection.

Paper Overview

Field: Machine Learning
Authors: Yury Gorishniy, Ivan Rubachev, Dmitrii Feoktistov
Published: 2025-04-17
arXiv: 2504.13081

Key Findings

MLPs are a heavily used backbone in modern deep learning architectures for supervised learning on tabular data, and AdamW is the go-to optimizer for training tabular DL models. However, unlike architecture design, the choice of optimizer for tabular deep learning has not been systematically examined, despite new optimizers showing promise in other domains. This paper fills that gap.

The authors benchmark multiple optimizers on multiple tabular datasets for training MLP-based models in the standard supervised learning setting, using a shared experiment protocol.

Main Results

1. Muon consistently outperforms AdamW. The Muon optimizer should be considered a strong and practical choice for practitioners and researchers, provided the associated training efficiency overhead is affordable. 2. EMA of model weights helps. Exponential moving average (EMA) of model weights is a simple yet effective technique that improves AdamW's performance on plain MLPs, though its effect is less consistent across other model variants.

Original Abstract

> MLP is a heavily used backbone in modern deep learning (DL) architectures for supervised learning on tabular data, and AdamW is the go-to optimizer used to train tabular DL models. Unlike architecture design, however, the choice of optimizer for tabular DL has not been examined systematically, despite new optimizers showing promise in other domains. To fill this gap, we benchmark N optimizers on N tabular datasets for training MLP-based models in the standard supervised learning setting under a shared experiment protocol. Our main finding is that the Muon optimizer consistently outperforms AdamW, and thus should be considered a strong and practical choice for practitioners and researchers, if the associated training efficiency overhead is affordable. Additionally, we find exponential moving average of model weights to be a simple and effective technique for improving AdamW performance on plain MLPs.

---

*Auto-collected on 2026-04-18*