Jiawei Wang

Also known as JarvisUSTC

Jiawei Wang is an AI researcher and the Chief Scientist of SimpleAI in Beijing, where he heads the pretraining of embodied foundation models. Trained in the School of the Gifted Young at the University of Science and Technology of China (USTC), he earned a PhD in June 2026 through its joint program with Microsoft Research Asia (MSRA). His Detect-Order-Construct method for hierarchical document structure analysis became the basis for a feature that Azure AI Document Intelligence added in 2024. As an intern on DeepSeek's Multimodal LLM Team in 2024, he contributed to DeepSeek-VL2 and DeepSeek-V3, and DeepSeek-R1 lists him among its authors. At ByteDance Seed from 2025, he coauthored the WideSearch benchmark and the EMPG method for training long-horizon LLM agents before joining SimpleAI in July 2026.

Now at: SimpleAI
Badges: Bridge
Links: Google Scholar, X, LinkedIn and GitHub

Timeline

2017–2021

Gifted Young program at USTC

He studies computer science in the School of the Gifted Young at the University of Science and Technology of China (USTC) from 2017 and earns his B.S. in 2021. In 2020, he begins a research internship in the Multimodal Interaction Group at Microsoft Research Asia (MSRA) in Beijing.
2021–2024

Document AI at Microsoft Research Asia

In 2021, he enters the joint PhD program of USTC and MSRA, co-supervised by Qiang Huo at MSRA and Jun Du at USTC. With colleagues, he releases Detect-Order-Construct in January 2024, a tree-construction approach to hierarchical document structure analysis that comes with a new benchmark, Comp-HRDoc. In March 2024, Azure AI Document Intelligence adds hierarchical document structure analysis based on the paper.
2024

AGI research intern at DeepSeek

From June 2024 to August 2024, he is an AGI research intern on the Multimodal LLM Team at DeepSeek in Beijing. He contributes to DeepSeek-VL2, a series of mixture-of-experts vision-language models released in December 2024 in three variants with 1.0B, 2.8B and 4.5B activated parameters.
December 2024–January 2025

DeepSeek-V3 and DeepSeek-R1

In December 2024, the DeepSeek-V3 technical report credits him among its contributors; the model is a mixture-of-experts language model with 671B total parameters, 37B of them activated for each token. In January 2025, he is also listed among the authors of DeepSeek-R1, which shows that pure reinforcement learning can incentivize reasoning abilities in large language models.
2025

Defending vision-language models

At ICCV 2025, he and his coauthors show that many vision-language models are susceptible to simple perturbations such as Gaussian noise. They propose Robust-VLGuard, a multimodal safety dataset for noise-augmented fine-tuning, and DiffPure-VLM, a defense that uses diffusion models to convert adversarial perturbations into Gaussian-like noise.
April 2025–March 2026

Agent research at ByteDance Seed

From April 2025 to March 2026, he is a research intern on ByteDance's Seed team, where he works on LLM and multimodal LLM agent projects. He coauthors WideSearch in August 2025, a benchmark of 200 questions that tests how reliably search agents gather large-scale information. In September 2025, he and colleagues introduce Entropy-Modulated Policy Gradients (EMPG) for training long-horizon LLM agents, and ICML 2026 accepts the paper.
June 2026

A PhD from USTC

He receives his PhD in engineering from USTC in June 2026 through the joint program with MSRA. His doctoral research covers document intelligence, including OCR and document layout analysis, and large language models, including multimodal models, agents and retrieval-augmented generation.
Since July 2026

Chief scientist at SimpleAI

In July 2026, he joins SimpleAI in Beijing as Chief Scientist, heading the pretraining of embodied foundation models. He and SimpleAI colleagues release HiFi-UMI in July 2026, a portable system for collecting robot-free manipulation data; a policy post-trained only on its demonstrations deploys directly on a real robot.