Daya Guo researches language models, code intelligence and reasoning. At DeepSeek from July 2023 to April 2026, he contributed to its coding, mathematics and reasoning models, including DeepSeek-Coder and DeepSeek-R1.[1]
Start with DeepSeek-Coder →
GitHub Background
Guo earned a computer-science-and-technology bachelor’s degree and PhD at Sun Yat-sen University. His doctoral work was part of a joint program with Microsoft Research Asia, where he also held research internships. His research asks how models can understand and generate both natural language and programming languages, connecting code structure with the statistical learning methods used for text.[1]
Career
His Microsoft Research Asia internships ran from July 2017 to May 2023, first with Duyu Tang and later with Nan Duan. He then worked as an AI researcher at DeepSeek, focusing on code intelligence and language-model reasoning through April 2026.[1]
Contributions
Training models for software
Guo is first author of the DeepSeek-Coder paper. The team trained code models from scratch on a mixture dominated by source code, with additional natural-language material. Project-level data and a fill-in-the-middle objective support completing code with surrounding context, rather than only continuing a short prefix. Instruction tuning then adapts the models to requests expressed in ordinary language. The released family spans different model sizes.[2]
Mathematics and verifiable reasoning
He coauthored DeepSeekMath, which combines continued training on mathematical material with reinforcement learning. Its Group Relative Policy Optimization compares sampled responses within a group, avoiding the separate critic model used by PPO. He also contributed to DeepSeek-R1. That report distinguishes an experiment that applies reinforcement learning without preliminary supervised fine-tuning from a later training pipeline that adds cold-start data and multiple stages. These are team efforts in model training and evaluation.[3][4]
Tools, writing & talks
Writing, research & talks
Start here
DeepSeek-Coderarxiv.orgDescribes a code-focused training corpus, model family and evaluation across programming tasks.
Read the data and training-objective sections to see what changes when a language model is trained for code. Basic language-model knowledge helps; the abstract provides an accessible map before the benchmark tables.
First author[2]- DeepSeekMatharxiv.org
Explains mathematical pretraining and Group Relative Policy Optimization.
Coauthor[3] - DeepSeek-R1arxiv.org
A technical report on reinforcement learning and multistage reasoning-model training.