Lei Cui

Also known as 崔磊

Lei Cui is a Senior Principal Research Manager in the General Artificial Intelligence group at Microsoft Research Asia in Beijing, and his document AI work covers models and datasets for reading scanned forms, receipts and tables. He began at the lab as an intern in 2007, earned a PhD from the Department of Computer Science at Harbin Institute of Technology, and joined Microsoft Research Asia in January 2015. With colleagues, he coauthored LayoutLM, which jointly models text and page layout and appeared at KDD 2020, along with the TableBank and DocBank datasets, the TrOCR text recognizer and LayoutLMv3. His later papers moved into multimodal and generative models, including TextDiffuser and KOSMOS-2.5 in 2023, and in 2025 he coauthored Geometric-Mean Policy Optimization for training reasoning language models.

Now at: Microsoft Research Asia
Works: 79
Citations: 2,566
h-index: 20
Badges: Bridge
Wayfarer

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2007–2014

Intern and joint PhD student

He begins an internship at Microsoft Research Asia in 2007. As a doctoral student trained jointly by Harbin Institute of Technology's Ministry of Education–Microsoft Key Laboratory of Language and Speech and Microsoft Research Asia, he coauthors “Learning Topic Representation for SMT with Neural Networks” at ACL 2014.
January 2015

Joining Microsoft Research Asia

After earning a PhD from the Department of Computer Science at Harbin Institute of Technology, he joins Microsoft Research Asia in January 2015.
2019–2020

TableBank and DocBank datasets

Starting in 2019, he and colleagues release TableBank, with 417K labeled tables for table detection and recognition, and DocBank, with 500K document pages annotated at the token level for layout analysis. The two datasets appear at LREC and COLING in 2020.
2019–2020

LayoutLM for document images

In December 2019, he coauthors LayoutLM, a pre-trained model that jointly learns text and layout across scanned document images. It improves results on form understanding, receipt understanding and document image classification, and the paper appears at KDD 2020.
2021

TrOCR and an ICDAR talk

In 2021, he coauthors TrOCR, an end-to-end text recognition model built from pre-trained image and text Transformers. At an ICDAR 2021 workshop in September 2021, he presents Document AI benchmarks, models and applications, including LayoutLMv2 and the multilingual LayoutXLM.
2022

LayoutLMv3 with unified masking

He coauthors LayoutLMv3 in 2022, which pre-trains multimodal Transformers for Document AI with unified text and image masking and a word-patch alignment objective. The paper appears at ACM Multimedia 2022.
2023

TextDiffuser and KOSMOS-2.5

In 2023, he coauthors TextDiffuser, a NeurIPS 2023 paper on diffusion models that render coherent text inside images, released with the MARIO-10M dataset of 10 million image-text pairs. KOSMOS-2.5, submitted in September 2023, is a multimodal literate model that reads text-intensive images and outputs spatially aware text blocks or markdown.
As of October 2026

Research manager in General AI

He is a Senior Principal Research Manager in the General Artificial Intelligence group at Microsoft Research Asia in Beijing, with research interests in foundation models and multimodal AI. In 2025, he coauthors Geometric-Mean Policy Optimization, which steadies GRPO training of reasoning language models by maximizing the geometric mean of token-level rewards. His record includes more than 60 papers and more than 20 US patents in AI and NLP.