Mao Yang

Also known as 杨懋

Mao Yang is a computer-systems researcher whose work at Microsoft Research Asia (MSRA) connects distributed systems with search, retrieval and language models. After computer-science degrees at the Harbin Institute of Technology and a PhD at Peking University in 2006, they joined MSRA in 2006 and worked on the Kirin, Tiger and Cougar systems behind Bing. Their co-authored papers include SPANN in 2021, the NeurIPS 2022 Outstanding Paper A Neural Corpus Indexer for Document Retrieval, LongRoPE and T-MAC in 2024, and rStar-Math in 2025. As of 2025, they are an assistant managing director of MSRA, focusing on AI infrastructure and tools and on algorithms for web search.

Now at: Microsoft Research Asia
Badges: Bridge

Timeline

2000–2006

Harbin and Peking University degrees

Yang earns a bachelor's degree in 2000 and a master's degree in 2002 in computer science at the Harbin Institute of Technology. A PhD in computer science from Peking University follows in 2006.
2006–2010

Joining MSRA and building Kirin

Yang joins Microsoft Research Asia in Beijing in 2006 and serves as research manager for its Systems and Networking Research Group (Asia). On the Bing team, they design and implement the replication and failover protocol of Kirin, a web store and processing system that starts processing many billions of web data in 2010.
2012–2016

Tiger, Cougar and Bing answers

As an architect on the Bing team, Yang works on Tiger, a flash-memory-based index serving platform that starts serving all web queries in 2012, and Cougar, a ranking system for semantic ranking models that starts serving all web queries in 2013. Their web-scale question-answering system starts giving direct answers in Bing in 2016.
2021

SPANN for billion-scale search

In 2021, Yang co-authors SPANN, a memory-disk hybrid system for approximate nearest neighbor search that keeps cluster centroids in memory and large posting lists on disk. Presented as a NeurIPS 2021 spotlight, SPANN reaches 90% recall twice as fast as DiskANN at the same memory cost on three billion-scale datasets.
2022

A NeurIPS Outstanding Paper

A Neural Corpus Indexer for Document Retrieval, co-authored by Yang with colleagues from Microsoft, Tsinghua University, the University of Illinois and Peking University, receives a NeurIPS 2022 Outstanding Paper award. Its Neural Corpus Indexer is a sequence-to-sequence network that generates document identifiers directly for a query, and it improves Recall@1 on NQ320k by 21.4% and R-Precision on TriviaQA by 16.8% over the best baseline.
2024

LongRoPE and T-MAC

LongRoPE, an ICML 2024 paper co-authored by Yang, extends the context window of pre-trained language models to 2048k tokens and is integrated into Microsoft Phi-3. In 2024, the lookup-table method T-MAC, also co-authored by Yang, runs low-bit language models on CPUs with up to 4 times the throughput and 70% lower energy use than llama.cpp.
2025

rStar-Math for small models

rStar-Math, an ICML 2025 oral paper co-authored by Yang, shows that small language models can rival OpenAI o1 in math reasoning without distillation from superior models. Using Monte Carlo Tree Search guided by a process reward model, it lifts Qwen2.5-Math-7B from 58.8% to 90.0% on the MATH benchmark.
By 2025

Assistant managing director at MSRA

By 2025, Yang is assistant managing director of Microsoft Research Asia, with research interests in distributed systems, information retrieval systems and machine learning systems. Their current focus is AI infrastructure and tools and algorithms for web search, and the projects they release include OpenPAI, NNI and MMdnn.