Yize Wu (吴亦泽)

@ PhD candidate at ISCAS (Institute of Software Chinese Academy of Sciences)

I am interested in Machine Learning Systems and AI Infra for LLMs, especially in multi-GPU distributed training/inference. I am also focusing on the engineering work that turns research ideas into dependable system software.

Email me GitHub Publications

News

2026.4: One first-author work EasyBalance: Cross-Layer Load Balancing in Distributed MoE Inference is accepted at ICML 2026. See you in Seoul!

2026.4: Our work Beyond Surface Level Pattern Trap: LLM Agents for Faster and Smarter Cross-Architecture Code Migration is accepted at ACL 2026 as Findings. See you in San Diego!

2026.1: One first-author work Stable-LoRA: Stabilizing Feature Learning of Low-Rank Adaptation is accepted at ICLR 2026. See you in Rio de Janeiro!

2025.9: One first-author work EasySpec: Layer-Parallel Speculative Decoding for Efficient Multi-GPU Utilization is accepted at NeurIPS 2025. See you in San Diego!

Publications

EasySpec: Layer-Parallel Speculative Decoding for Efficient Multi-GPU Utilization

Yize Wu, Ke Gao, and Yanjun Wu

A layer-parallel speculative decoding strategy that improves GPU utilization during draft-model inference while preserving the base model distribution.

arXiv Code
LLM inference speculative decoding multi-GPU

Stable-LoRA: Stabilizing Feature Learning of Low-Rank Adaptation

Yize Wu, Ke Gao, Ling Li, and Yanjun Wu

A weight-shrinkage optimization strategy for LoRA that improves feature learning stability during early training with negligible additional overheads.

arXiv Code
LoRA feature learning LLM fine-tuning

EasyBalance: Cross-Layer Load Balancing in Distributed MoE Inference

Yize Wu, Ke Gao, Ling Li, and Yanjun Wu

A cross-layer MoE load balancing strategy that improves GPU utilization without requiring expert replication or migration.

arXiv Code
LLM Inference expert parallelism MoE

Repos

EasyInfra

Code for papers in the EasyInfra series. (Including: EasySpec, EasyBalance)

LLM inference

Stable-LoRA

Implementation of Stable-LoRA, a weight-shrinkage optimization strategy for stabilizing LoRA feature learning.

LoRA Fine-tuning

Educations

PhD Candidate in Software Engineering — Institute of Software, Chinese Academy of Sciences, Beijing

2022 — present. Advisor: Professor Yanjun Wu.

Focus: Machine Learning Systems and AI infrastructure.

BS in Computer Sciences and Technology — University of Chinese Academy of Sciences, Beijing

2018 — 2022. Tutor: Professor Huimin Lin.

Social-media Links