Shuai Zhang (@DavenCheung) header
Shuai Zhang (@DavenCheung) profile picture

Now senior applied scientist @amazon, @awscloud. Previously, Postdoc @ETH zurich. Opinions are my own.

Santa Clara
Joined August 2016
Shuai Zhang retweeted
We built AMMO to let AI agents autonomously optimize vLLM. - No human chose the bottleneck. - No human proposed the optimization. - No human wrote or tuned the mechanism. - No human decided whether it passed the performance gates. AMMO handled: profiling → implementation → integration → verification It produced the changes behind 4 public vLLM PRs. And we open-sourced the whole system so anyone can run it. Getting there exposed what I think is a much broader problem with long-horizon agents. 🧵
5
2
3
3
715
I am happy to share that our work on efficient RL training for large rollouts and long conetxts is accepted by NeurIPS 2026 #NeurIPS2026 #RL. Check out the code here: github.com/amazon-science/du…
2
28
1,750
FULL GUIDE: HOW TO SCALE YOUR MODEL: jax-ml.github.io/scaling-boo… PART 12: HOW TO THINK ABOUT GPUS: jax-ml.github.io/scaling-boo… I have one more thing before you go. If you want to become a generative AI data scientist in 2026 ($200,000 career), then I'd like to help:
1
10
36
2,095
Excited to share that OPERA is accepted to COLM 2026! OPERA makes dense retriever adaptation more efficient by dynamically pruning training data at both query and document levels, achieving comparable performance in less than 50% of the training time. #COLM
1
2
19
1,601
RL training with large rollouts wastes massive compute as they share the same long prompt. We propose DualKV, a FlashAttention variant that removes redundant computation while preserving exact full attention. 3.82× policy-update and 3.38× end-to-end speedup at 30B MoE scale.
5
8
86
6,075
Paper: arxiv.org/pdf/2605.15422 Code: github.com/amazon-science/du… The code can be patched to veRL and flashattention to support production-grade RL training.
2
588
Excited to be in Bellevue next week for #MLSys2026 and presenting our work on HetRL! Happy to chat about RL systems, large-scale LLM training, etc. — feel free to reach out if you'll be around.
Excited to be giving the oral presentation for our paper "HetRL: Efficient Reinforcement Learning for LLMs in Heterogeneous Environments" at #MLSys2026 in Bellevue next Wednesday at 3:30PM! In this work, we study how to scale RL training for LLMs across heterogeneous accelerators. Huge thanks to my amazing co-authors at @AWSAI and EC2: Shuai Zhang @DavenCheung, Jiading Gai, Xiyuan Zhang @xiyuanzh, Boran Han @BoranHan1742511, Yuyang (Bernie) Wang, Huzefa Rangwala @DMfun, George Karypis @karypis. We're also grateful to other colleagues @AmazonScience for their invaluable feedback: Yida Wang, Annan Yu, Zelin He, Hengyi Wang, Danielle Maddix Robinson. This work would not have been possible without their contributions and encouragement. If you'll be in Bellevue for MLSys, happy to chat and grab a drink! Always open to discussions and potential collaborations! Paper: arxiv.org/abs/2512.12476 MLSys schedule: mlsys.org/virtual/2026/oral/…
2
5
757
Shuai Zhang retweeted
OPERA: Online Data Pruning for Efficient Retrieval Model Adaptation Amazon presents a data pruning framework for dense retriever finetuning that adaptively modulates sampling probabilities, improving both ranking and retrieval. 📝 arxiv.org/abs/2603.17205
3
16
981
Shuai Zhang retweeted
🚀 Qwen3.5-397B-A17B is here: The first open-weight model in the Qwen3.5 series. 🖼️Native multimodal. Trained for real-world agents. ✨Powered by hybrid linear attention + sparse MoE and large-scale RL environment scaling. ⚡8.6x–19.0x decoding throughput vs Qwen3-Max 🌍201 languages & dialects 📜Apache2.0 licensed 🔗Dive in: GitHub: github.com/QwenLM/Qwen3.5 Chat: chat.qwen.ai API:modelstudio.console.alibabac… Qwen Code: github.com/QwenLM/qwen-code Hugging Face: huggingface.co/collections/Q… ModelScope: modelscope.cn/collections/Qw… blog: qwen.ai/blog?id=qwen3.5
267
842
418
5,204
1,280,985
Excited to share that our paper HetRL has been accepted to #MLSys2026 🚀 We present a new system for scalable RL training of LLMs on heterogeneous GPU clusters, delivering up to 9.17× speedup over state-of-the-art approaches. Paper: arxiv.org/pdf/2512.12476 #MLSys2026 #LLM #RL
2
8
404
🚀 We’re hiring PhD Applied Scientist Interns (Summer 2026) to work on automated kernel optimization, leveraging reinforcement learning, multi-agent systems, and more! 💡 Apply here: amazon.jobs/en/jobs/3051006/ 📧 Send your resume to [email protected] #LLMs #LLM4Code #Internship
1
9
93
6,378
Mitra shows the power of synthetic pretraining — achieving SOTA on real tabular datasets and already hitting 3M+ downloads on HF. Paper: arxiv.org/pdf/2510.21204
🔥 Excited to share our NeurIPS paper: “Mitra: Mixed Synthetic Priors for Enhancing Tabular Foundation Models”! 🚀 Mitra is pretrained purely on synthetic data and archives SOTA on real datasets. 🤗 Mitra has received over 3 million downloads on HF in just 3 months of release!
4
558
I’m attending #KDD2025 this week in Toronto! 🇨🇦 We will present PipeRAG, rethinking RAG for systems efficiency—achieving 2.6× faster latency with: 🔁 Pipeline parallelism 📦 Stale-query prefetching 📈 Query-aware performance modeling 📄 Paper: arxiv.org/abs/2403.05676 #RAG #LLM
1
465
Can your RAG pipeline trust the documents it retrieves? We propose a method - CrEst - to estimate document credibility—without human labels—and offer a recipe to integrate it into open-source or private models. Check it out: arxiv.org/pdf/2506.14912
5
500
We should do the thinking, not LLMs.
BREAKING: MIT just completed the first brain scan study of ChatGPT users & the results are terrifying. Turns out, AI isn't making us more productive. It's making us cognitively bankrupt. Here's what 4 months of data revealed: (hint: we've been measuring productivity all wrong)
2
498
🚀 Excited to announce #GENNEXT at #SIGIR2025 — the first workshop dedicated to next-gen IR & recommender systems with #agents, LLMs, and conversational AI! 📄 Papers will be indexed in CEUR 🗓️ Submission deadline: April 23, 2025 sigirgennext.github.io/GENNE…
1
1
4
744
I'm excited to share that I've been promoted to Senior Applied Scientist at Amazon Web Services (AWS)!
5
1
67
7,058
🚀 We are coming back to @SIGIRConf ! Delighted to share that our workshop *GENNEXT: The Next Generation of IR and Recommender Systems with #LANGUAGE #Agents, Generative Models, and Conversational AI* has been accepted to #SIGIR2025. 📍GENNEXT: July 13-18,2025 (Padua, Italy) @ScottSanner @pcastells @enricopalumbo91 @huguesbouchard @JulianMcauley @DavenCheung
1
7
1,420