Nan Yang

Nan Yang is a Principal Researcher in the General Artificial Intelligence group at Microsoft Research Asia, working on language models and text embeddings. As a coauthor, they wrote the gated self-matching network for reading comprehension in 2017, the UniLM pre-trained language model in 2019 and the compressed Transformer MiniLM in 2020. In 2022, they joined Liang Wang, Furu Wei and colleagues on the E5 text embeddings, the first model to outperform BM25 on the BEIR retrieval benchmark without labeled data. Their 2023 and 2024 papers cover embeddings built with large language models and the multilingual E5 models, and their 2025 paper introduces Chain-of-Retrieval Augmented Generation.

Now at: Microsoft Research Asia
Works: 206
Badges: Bridge

Publication numbers from OpenAlex, as of 22 October 2026. Citations and h-index are hidden because this OpenAlex record may be incomplete or mix in other authors.

Timeline

2014

Sentiment-specific word embeddings

With Duyu Tang, Furu Wei, Ming Zhou, Ting Liu and Bing Qin, they coauthor 'Learning Sentiment-Specific Word Embedding for Twitter Sentiment Classification' in 2014. The paper appears in the proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics in Baltimore.
2017

Gated self-matching networks

The 2017 paper 'Gated Self-Matching Networks for Reading Comprehension and Question Answering', written with Wenhui Wang, Furu Wei, Baobao Chang and Ming Zhou, appears at the ACL meeting in Vancouver. Its single model reaches 71.3% exact match on the hidden SQuAD test set, the ensemble reaches 75.9%, and the model holds first place on the SQuAD leaderboard at the time of submission.
2019

The UniLM language model

Their 2019 paper 'Unified Language Model Pre-training for Natural Language Understanding and Generation' presents UniLM, written with Li Dong, Wenhui Wang, Furu Wei and others. One pre-trained model with a shared Transformer network is fine-tuned for both understanding and generation tasks, and it sets new state-of-the-art results on five natural language generation datasets. NeurIPS 2019 accepts the paper.
2020

Compressing models with MiniLM

MiniLM, coauthored in 2020 with Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao and Ming Zhou, compresses large pre-trained Transformer models. A small student model learns by distilling the self-attention module of the last Transformer layer of the teacher.
2022

The E5 text embeddings

In 2022, they coauthor 'Text Embeddings by Weakly-Supervised Contrastive Pre-training' with Liang Wang, Furu Wei and others, which introduces the E5 text embeddings. E5 is trained on the CCPairs text-pair dataset and is the first model to outperform BM25 on the BEIR retrieval benchmark without labeled data; fine-tuned, it obtains the best results on the MTEB benchmark.
2023–2024

Multilingual E5 and LLM embeddings

The 2023 paper 'Improving Text Embeddings with Large Language Models', written with Liang Wang, Furu Wei and others, has proprietary LLMs generate synthetic data for hundreds of thousands of embedding tasks across 93 languages, and ACL 2024 accepts it. In 2024, they also coauthor the technical report on the open-source multilingual E5 models, released in mid-2023.
2025

Chain-of-retrieval augmented generation

Chain-of-Retrieval Augmented Generation (CoRAG), written with Liang Wang, Haonan Chen, Xiaolong Huang, Zhicheng Dou and Furu Wei, appears in January 2025. It trains o1-like RAG models that retrieve and reason step by step before generating the final answer, and NeurIPS 2025 accepts the paper.
As of October 2026

Principal researcher in General AI

As of October 2026, they are a Principal Researcher in the General Artificial Intelligence group at Microsoft Research Asia, with human language technologies as their research area.