Hao Zhang
Assistant Professor
Builds systems for training and serving large machine learning models, such as vLLM and distributed LLM serving.
Hao Zhang's lab at UCSD works on efficient training & inference, distributed, cloud & edge systems and data management & big data. 35 papers since 2023, 31 at top venues, mostly ICLR and ICML. 8 PhD students and 2 master's students. 25 alumni, most often to ByteDance and NVIDIA. Applications close .
At a glance
- papers, 2023–2026
- 9/ year
- at T1 venues
- 89%
- PhD students
- 8
- MS students
- 2
- alumni listed
- 25
Papers by yearfrom the lab homepage, 2023–2026
- 5 papers2026 · T1 5 · collecting
- 11 papers2025 · T1 9
- 13 papers2024 · T1 12
- 6 papers2023 · T1 5
T1 venuesOtherVenues: ICLR, ICML, NeurIPS, ACL, MLSys
Research focuspaper titles that use each topic's words
- Efficient training & inference13 papers
- serving4 papers
- Data management & big data2 papers
- Distributed, cloud & edge systems1 paper
- serving large0 papers
Paper list
From the lab homepage.
20265 papers
- ICLR (T1)Deep Think with Confidence (opens in a new tab)
- ICLR (T1)Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcing (opens in a new tab)
- MLSys (T1)Efficient Long-context Language Model Training by Core Attention Disaggregation (opens in a new tab)
- ICLR (T1)LMGame-Bench: How Good are LLMs at Playing Games? (opens in a new tab)
- ICLR (T1)Stronger Together: On-Policy Reinforcement Learning for Collaborative LLMs (opens in a new tab)
202511 papers
- ES-FOMO Workshop @ ICMLEfficient-vDiT: Efficient Video Diffusion Transformers With Attention Tile (opens in a new tab)
- NeurIPS (T1)Efficiently Scaling LLM Reasoning with Certaindex (opens in a new tab)
- ICML (T1)Fast Video Generation with Sliding Tile Attention (opens in a new tab)
- ICLR (T1)GameArena: Evaluating LLM Reasoning through Live Computer Games (opens in a new tab)
- ICLR (T1)MatryoshkaKV: Adaptive KV Compression via Trainable Orthogonal Projection (opens in a new tab)
- TPAMI (T1)Neural Eigenfunctions Are Structured Representation Learners (opens in a new tab)
- Foundation Models in the Wild Workshop @ ICMLReasoning without Self-doubt: More Efficient Chain-of-thought through Certainty Probing (opens in a new tab)
- ICLR (T1)Scaling Long Context Training Data by Long-Distance Referrals (opens in a new tab)
- NeurIPS (T1)Scaling Speculative Decoding with Lookahead Reasoning (opens in a new tab)
- ACL (T1)TrimLLM: Progressive Layer Dropping for Domain-Specific LLMs (opens in a new tab)
- NeurIPS (T1)VSA: Faster Video Diffusion with Trainable Sparse Attention (opens in a new tab)
202413 papers
- EMNLP (T1)AdaMoE: Token-Adaptive Routing with Null Experts for Mixture-of-Experts Language Models (opens in a new tab)
- ICML (T1)Break the Sequential Dependency of LLM Inference using Lookahead Decoding (opens in a new tab)
- ICML (T1)Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (opens in a new tab)
- ICML (T1)CLLMs: Consistency Large Language Models (opens in a new tab)
- COLMDistFlashAttn: Distributed Memory-efficient Attention for Long-context LLMs Training (opens in a new tab)
- OSDI (T1)DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving (opens in a new tab)
- ACL (T1)Efficient Detection of LLM-generated Texts with a Bayesian Surrogate Model (opens in a new tab)
- NeurIPS (T1)Efficient LLM Scheduling by Learning to Rank (opens in a new tab)
- ICML (T1)InferCept: Efficient Intercept Support for Augmented Large Language Model Inference (opens in a new tab)
- ICLR (T1)LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset (opens in a new tab)
- NeurIPS (T1)Megalodon: Efficient LLM Pretraining and Inference with Unlimited Context Length (opens in a new tab)
- ICML (T1)MuxServe: Flexible Spatial-Temporal Multiplexing for Multiple LLM Serving (opens in a new tab)
20236 papers
- OSDI (T1)AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving (opens in a new tab)
- SOSP (T1)Efficient Memory Management for Large Language Model Serving with PagedAttention
- Instruction Tuning and Instruction Following Workshop @ NeurIPSHow Long Can Context Length of Open-Source LLMs truly Promise? (opens in a new tab)
- NeurIPS (T1)Judging LLM-as-a-judge with MT-Bench and Chatbot Arena (opens in a new tab)
- ICLR (T1)MPCFormer: Fast, Performant and Private Transformer Inference with MPC (opens in a new tab)
- MLSys (T1)On Optimizing the Communication of Model Parallelism (opens in a new tab)
Where alumni went25 listed, 59% to industry
Industry 13Academia 9
Most often to
- ByteDance
- NVIDIA
- Alibaba
Also to
- Carnegie Mellon University
- Harvard University
- NUS
Apply to UCSDFall 2027
UCSD CSE Computer Science PhD
- Round
- Single round
- Opens
- —
- GRE
- See program page
- English
- See program page
- Fee
- See program page
Not yet confirmed for this intake.
Similar labsnearest by research summary
- CHETH Zurich
Ana Klimovic
- Efficient training & inference
- USUT Austin
Neeraja Yadwadkar
- Efficient training & inference
- USUSC
Konstantinos Psounis
- Efficient training & inference
- CNSJTU
Xingda Wei
- Efficient training & inference
- CNSJTU
Chao Li
- Accelerators, chips & architecture
- CNSJTU
Qinya Li
- Distributed, cloud & edge systems
Your notes
Save this lab to keep notes.
Data
Checked against the lab homepage.
- Collected
- Verified
- Papers
- The lab homepage
- Members
- Lab homepage
- Alumni
- Lab homepage
Tiers are venue classifications, not lab rankings.
UCSDcloses