Hao-Cong Wu | Academic Homepage

About Me

I am an undergraduate student in the School of Artificial Intelligence at Nanjing University, expected to join the LAMDA Lab as a graduate student under the supervision of Prof. Peng Zhao. My research centers on efficient inference for large language models, with a particular focus on speculative decoding, online learning, and diffusion language models. I am also broadly interested in LLM post-training and scalable inference systems. I currently work as an inference optimization intern at StepFun, focusing on efficient distributed training and deployment of speculative-decoding draft models.

Highlights

[2026] One paper accepted to ICML 2026 🎉
[2026] Published AdaFlash: Adaptive Speculative Decoding via On-Policy Distilled Diffusion Drafters on arXiv
[2026] Joined StepFun as an LLM inference optimization intern

Research Interests

LLM Efficiency — accelerating autoregressive inference via parallel and speculative decoding.
Speculative Decoding — evolving draft models with online learning and on-policy distillation.
Diffusion Language Models — discrete masked / bidirectional diffusion LLMs as efficient drafters.
Online Learning & LLM Post-Training — adapting models on-the-fly during inference.

Preprints

AdaFlash: Adaptive Speculative Decoding via On-Policy Distilled Diffusion Drafters arXiv Code Model

Yu-Yang Qian*, Hao-Cong Wu*, Chen Chen, Peng Zhao, and Zhi-Hua Zhou.

Co-first authors.

Publications

When Drafts Evolve: Speculative Decoding Meets Online Learning Paper Code

Yu-Yang Qian, Hao-Cong Wu, Yichao Fu, Hao Zhang, Peng Zhao
ICML 2026 — Proceedings of the 43rd International Conference on Machine Learning.

Applies online learning to speculative decoding, continuously updating the draft model during inference to improve efficiency. The proposed OnlineSPEC achieves up to 24% speedup over state-of-the-art.

Projects

Agentic Inference Acceleration via Online Learning Huawei Collaboration · Apr–Aug 2026

Core student contributor to a joint project between the Nanjing University–Huawei Joint Lab, applying speculative-decoding research to Huawei Agent inference.
Implemented vertical-domain on-policy distillation (OPD) fine-tuning with deployment-time draft-model updates, and dynamically adjusted the verification length under high-concurrency scenarios to boost throughput.
A patent application has been filed, with planned deployment in Huawei terminal Agent scenarios.

StepFun · LLM Inference Optimization Engineer Internship · Jul 2026 – Present

Working on inference acceleration for next-generation foundation models, including efficient distributed training and algorithmic deployment of speculative-decoding draft models.

SGLang × Diffusion LLM Open Source · PR #20615

d3LLM (arXiv:2601.07568, GitHub 146 ⭐) is a discrete masked diffusion language model with bidirectional attention, enabling parallel-decoding acceleration.
Adapted the d3LLM model family into SGLang and implemented its inference interface (PR #20615).
Achieved 1300 tokens/s inference speed on an NVIDIA B200, a 5× speedup over autoregressive models.

Education

2023.09 – Present Nanjing University — B.Eng. in Artificial Intelligence, School of Artificial Intelligence

Skills

Programming: Python, PyTorch — deep learning development.
Systems: LLM inference engines, eg. SGLang, vLLM.
Algorithms: Reproduced speculative decoding methods including EAGLE, DFlash, and JetSpec.
Development tools: Claude Code, Codex, Cursor.

Python PyTorch Linux Speculative Decoding Diffusion LLM Online Learning SGLang LLM Inference