Liliang Ren
Member of Technical Staff at Thinking Machines Lab; first author of SambaY and Samba
- Member of Technical Staff, Thinking Machines Lab1
Liliang Ren is a Member of Technical Staff at Thinking Machines Lab and was previously at Microsoft Superintelligence, according to Ren's site. Ren is the first author of "Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation" (2025), which introduced the SambaY architecture used in Microsoft's Phi-4-mini-flash-reasoning model, and of Samba (2024), a hybrid state-space architecture for language modeling with unlimited context. Ren earned a PhD in computer science at the University of Illinois Urbana-Champaign, advised by ChengXiang Zhai.123
Career
Ren completed a PhD in computer science at the University of Illinois Urbana-Champaign, advised by ChengXiang Zhai, after a master's degree at the University of California San Diego. During the PhD, Ren held research internships at Microsoft Research in 2022 and at Amazon Science in 2023.1
Ren later worked at Microsoft Superintelligence and is now a Member of Technical Staff at Thinking Machines Lab.1
Research
Ren's research aims at neural architectures with greater efficiency and capacity. Samba, posted in June 2024 with Weizhu Chen as senior author, combines state-space layers with sliding-window attention for language modeling with unlimited context length. The SambaY paper, posted in July 2025, introduces a decoder-hybrid-decoder design and was released with Phi-4-mini-flash-reasoning and a training codebase.132
Ren's site notes five papers accepted at NeurIPS 2025, including SambaY and Routing Mamba, and a March 2026 paper on hyperparameter transfer for scaling training compute.1
Work
Phi-4-mini-flash-reasoningOpen weights
First author of the SambaY paper behind Phi-4-mini-flash-reasoning; Ren's site announces the model's July 2025 release together with the paper and its training codebase.214
Samba (external site: arxiv.org)Published research
First author of Samba, a hybrid architecture combining state-space layers with attention for efficient long-context language modeling, presented at ICLR 2025.31
Career timeline
May 2022
Began a research internship in the Knowledge and Language Team at Microsoft Research.1
Jun 2024
Posted the Samba paper on hybrid state-space language models as first author.3
Jul 2025
Released Phi-4-mini-flash-reasoning with the SambaY paper as first author.12
Sep 2025
Had five papers accepted at NeurIPS 2025, including SambaY.1
Mar 2026
Announced a paper on stable and optimal learning-rate transfer for training compute scaling.1
Other documented roles
Roles the sources document but that are not confirmed as current.
- Former affiliation, listed on Ren's site (past), Microsoft Superintelligence1
Related people
More sourced profiles at the same organization, in the same field, or from the same school.
- Alexander KirillovAlso at Thinking Machines Lab
- Akshita BhagiaLanguage models
- Ali FarhadiAlso from UIUC
- John SchulmanAlso at Thinking Machines Lab
- Arthur MenschLanguage models
- Arvind KrishnaAlso from UIUC
- Kenneth LiAlso at Thinking Machines Lab
- Barret ZophLanguage models
- Chris LattnerAlso from UIUC
- Mira MuratiAlso at Thinking Machines Lab
- Beren MillidgeLanguage models
- David ForsythAlso from UIUC
Is this you?
Profiles cover public professional work only, and every line cites a source. If something is out of date, missing or wrong, send a correction with a link to an official page and we will update it. Suggest a correction
Add this badge to your website or GitHub profile README. It links back to this page. It is not an endorsement or a rating.
Markdown (for a README)
[](https://unitedstatesofamericasuperintelligence.com/people/liliang-ren/)
HTML
<a href="https://unitedstatesofamericasuperintelligence.com/people/liliang-ren/"><img src="https://unitedstatesofamericasuperintelligence.com/badges/profile.svg" alt="Liliang Ren — Profile on USASI" height="20"></a>
This profile covers public professional work only, from the sources below. It is not an endorsement or a ranking. To correct or remove it, use the correction form.