Dacheng Li

I am a second-year CS PhD at EECS, UC Berkeley, fortunately advised by Prof. Ion Stoica and Prof. Joseph Gonzalez . I also have the fortunate to work closely with Prof. Song Han (MIT). I am affliated with Sky Computing Lab and Berkeley AI Group (BAIR). I am co-leading NovaSky and the 1st new member of lmsys.

Previously, I obtained my master in Machine Learning at CMU with Prof. Eric Xing , where I also worked with Prof. Hao Zhang . I obtained my undergraduate with double majors in Computer Science and Mathematics at UC San Diego with Prof. Zhuowen Tu . I interned at Nvidia Research and Google.

Extended acknowledgment to Rulin Shao (my girlfriend, PhD at UW), Lianmin Zheng and Ying Sheng (Professors at UCLA).

I am open for collaboration and opportunity. Please contact me for presentations, research 1-1 and collaborations.

Google Scholar  /  GitHub  /  Resume  /  PhD SoP  /  Twitter  /  Linkedin

Machine Learning Systems (MLSys)
Distributed training systems:: AMP (Neurips'22), DistFlashAttn (COLM'24), LongVILA (ICLR'25), MCBench (MLsys'24) are distributed training systems for LLMs under heterogeneous hardware, long context, visual-language models and model parallelism compression.
Distributed serving systems:: MPCFormer (ICLR'23) and Marill (Arxiv'24) accelerates the inference of LLMs for private and secure inference. VTC (OSDI'24) and DLPM (Arxiv'25) are distributed systems for LLM fair scheduling. S-lora (MLsys'24) is a distributed serving system for LoRA.

Machine Learning (ML)
Large Language Models: LLM-as-a-judge (Neurips'23) is one of the foundations of LLMs applications. LongChat (Neurips'23 workshop) is a long-context LLM.
Large Reasoning Models: Sky-T1 (Arxiv'25) is an open large reasoning model with o1 performance trained within academic budget. S* (Arxiv'25) is a simple and extensible test time scaling framework for code generation. Overthinking (Arxiv'25) reveals the Reasoning-Action Dilemma in Agentic Tasks.
Computer Vision: DC-VAE (CVPR'21) is the first VAE that can perform image synthesis at GAN level. Efficient-Vdit (Arxiv'25) and SparseVideoGen (Arxiv'25) are efficient video generation models.
Visual Language Models: VILA-U (ICLR'25) and NVILA (Arxiv'25) are efficient visual language models.
Evaluation: Chatbot Arena (ICML'24) is one of the most influential LLM chatbot evaluation platform. LongEval (Neurips'23 workshop) is a long-context LLM evaluation platform. Sorry-Bench (ICLR'25) is a systematic evaluation of LLM refusal.

• Nvidia Graduate Fellowship Finalist, 2025-2026
• AMD AI & HPC Cluster Award, 2025
• Amazon Research Awards, 2022

• Sky-T1 covered by New York Times, Wall Street Journal, and The Information.

Beyond research collaboration, I commit seriously to mentoring junior students. If you are serious about working with me, please fill out this form. We will have regular meetings and we will succeeed.