Yi Liu (刘毅)

I am a Principal Architect at Baidu, leading a pre-training data team for the ERNIE large language model, with a focus on long-context data curation and strategy for pre-training and mid-training. Previously, I led the large model team at PaddlePaddle, working across infrastructure, development tools, algorithms, and applications for the ERNIE model series.

I have led over 10 open-source projects at Baidu that have collectively earned more than 200,000 GitHub stars, including PaddleOCR, ERNIE, PaddleFormers, and PaddleX. These projects power critical AI applications worldwide, spanning multilingual document understanding, large language model training, and large-scale deployment.

My technical expertise spans large language model pre-training and data strategy, particularly for long-context modeling, as well as computer vision, vision-language models, and autonomous driving. I hold a Ph.D. from Nanyang Technological University, Singapore (2017) and a B.Eng. from Harbin Institute of Technology, China (2013). Before joining Baidu in 2018, I was a Data Scientist at HP Labs Singapore. At Baidu, I also collaborated extensively with the Apollo team on autonomous driving technologies, contributing to Robotaxi and AD 2.0.

• Jul 2026 Our three works, RISTER, Real5-OmniDocBench, and RT-DocLayout, have been accepted by ECCV 2026!
• Jun 2026 We open-sourced PP-OCRv6, a three-tier OCR model family (1.5M/7.7M/34.5M) that surpasses mainstream VLMs with major accuracy gains, unified support for 50 languages, and faster inference across CPU, Apple M4, and GPU deployments.
• May 2026 We open-sourced PaddleOCR-VL-1.6, achieving SOTA on OmniDocBench v1.6 with over 96.3% accuracy, significantly improving table, ancient document, rare character, and seal recognition while maintaining full compatibility with v1.5.
• May 2026 A collaboration with Renmin University of China on SUDER has been accepted by KDD 2026!
• May 2026 I gave a talk at VALSE 2026 and CCIG 2026 on "PaddleOCR Multimodal Document Intelligent Parsing".
• Mar 2026 Our two works, PP-OCRv5 and PaddleOCR-VL, have been accepted by CVPR 2026!
• Show more news ▼
  • Feb 2026 ERNIE 5.0 (2.4T-parameter unified multimodal model) released. My team developed its ultra-scale RL infrastructure for stable and efficient post-training with MoE architectures.
  • Jan 2026 We open-sourced PaddleOCR-VL-1.5: A 0.9B lightweight VLM delivering SOTA document parsing and text spotting, leading across six scenarios. It adds new text spotting and seal recognition, supports 111 languages.
  • Jan 2026 We released PaddleFormers 1.0: It supports training for LLMs and VLMs, with extreme performance optimizations for key models like DeepSeek-V3 and GLM-4.5-Air, significantly outperforming Megatron-LM in training speed.
  • Dec 2025 We officially launched PaddleOCR.com, offering free hands-on experience with advanced document parsing capabilities.
  • Nov 2025 We open-sourced ERNIE-4.5-VL-28B-A3B-Thinking: a lightweight VLM (3B active) matching flagship models, excelling in visual reasoning, STEM solving, visual grounding, and video understanding with tool utilization.
  • Oct 2025 We open-sourced PaddleOCR-VL: a SOTA 0.9B VLM for document parsing, supporting 109 languages and recognizing complex elements (text, tables, formulas, charts).
  • Sep 2025 We open-sourced ERNIE-4.5-21B-A3B-Thinking: an enhanced reasoning model with improved performance on logical reasoning, mathematics, science, coding, featuring tool usage and 128K long-context.
  • Aug 2025 I gave a talk at IJCAI Conference on "PaddlePaddle Deep Learning Framework and its Support for Large Model Training and Inference".
  • Jun 2025 We open-sourced ERNIE 4.5: a family of large-scale multimodal models with 10 variants, including MoE models (47B/3B active, 424B total) and dense models (0.3B).
  • May 2025 We open-sourced PaddleOCR 3.0: featuring PP-OCRv5 (high-accuracy text recognition), PP-StructureV3 (general-purpose document parsing), and PP-ChatOCRv4 (intelligent document understanding).