Zixiao Wang

Zixiao Wang, Yifei Shen, Huishuai Zhang

Many optimizers can be interpreted as steepest-descent methods under norm-induced geometries, and thus inherit corresponding implicit biases. We introduce \nameA{} (\fullname{}), which combines spectral control from orthogonalized update directions with $\ell_\infty$-style coordinate control from sign updates. \nameA{} forms a Lion-style momentum direction, approximately orthogonalizes it via a few Newton--Schulz iterations, and then applies an entrywise sign, providing an efficient approximation to taking a maximal step over the intersection of the spectral and $\ell_\infty$ constraint sets (a scaled Hadamard-like set for matrix parameters). Despite the strong nonlinearity of orthogonalization and sign, we prove convergence under a mild, empirically verified diagonal-isotropy assumption. Across large-scale language and vision training, including GPT-2 and Llama pretraining, SiT image pretraining, and supervised fine-tuning, \nameA{} matches or outperforms AdamW and Muon under comparable tuning while using only momentum-level optimizer state, and it mitigates optimizer mismatch when fine-tuning AdamW-pretrained checkpoints.

Zixiao Wang, Dong Qiao, Jicong Fan

The discrete distribution is often used to describe complex instances in machine learning, such as images, sequences, and documents. Traditionally, clustering of discrete distributions (D2C) has been approached using Wasserstein barycenter methods. These methods operate under the assumption that clusters can be well-represented by barycenters, which is seldom true in many real-world applications. Additionally, these methods are not scalable for large datasets due to the high computational cost of calculating Wasserstein barycenters. In this work, we explore the feasibility of using spectral clustering combined with distribution affinity measures (e.g., maximum mean discrepancy and Wasserstein distance) to cluster discrete distributions. We demonstrate that these methods can be more accurate and efficient than barycenter methods. To further enhance scalability, we propose using linear optimal transport to construct affinity matrices efficiently for large datasets. We provide theoretical guarantees for the success of our methods in clustering distributions. Experiments on both synthetic and real data show that our methods outperform existing baselines.