Jianbo Zhao

Master's Degree in Control Theory and Control Engineering

Institute of Automation, Chinese Academy of Sciences

State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS) · GPA: 3.84/4.00

Bachelor's Degree in Automation

Tianjin University

School of Electrical and Information Engineering · GPA: 3.89/4.00 · Top 3%

WeChat Input Engine Development

Tencent / WeChat Group / WeChat Keyboard Center

• Participated in the design of offline end-to-end models for WeChat keyboard, including structure optimization, data filtering, and KV cache acceleration, achieving a 4.1% improvement in Top1 accuracy compared to rule-based models.
• Led the knowledge distillation work for WeChat input method offline models, designing target-based decoupled knowledge distillation to accelerate soft logits learning, and improving the PKD method for better alignment. Introduced adversarial sample augmentation into the distillation framework, reducing model size by 60% with only a 1.7% performance decrease relative to the teacher model.

Spreadsheet Understanding and Language Model Training

Microsoft Research Asia / DKI Group / Excel Research

• Processed massive real-world spreadsheet data, including content extraction and data augmentation, to create datasets for fine-tuning spreadsheet-focused language models. Developed an end-to-end data processing pipeline that was adopted by MSRA and Excel Research teams and incorporated into the Excel Copilot codebase.
• Designed spreadsheet data encoding methods and trained spreadsheet language models (SpreadsheetLLM/SheetPacker), introducing (1) a novel approach for spreadsheet feature extraction and structured information representation, and (2) an efficient structured data compression method. The models achieved 89.4% F1 score (SOTA) on table boundary detection tasks and demonstrated strong generalization capabilities across multiple downstream tasks. Published three first/co-first author papers and one fourth-author paper.
• Collaborated with the Excel product team to implement SpreadsheetLLM in Excel Copilot-TableSense, which will be launched in the new Excel version this year.
• Personally received Microsoft Research Asia's "Rising Star" award (highest honor), and the team won the 2023-24 Microsoft AI Innovation Award.

CARES Embodied Intelligence Surgical Large Model

Hong Kong Institute of Artificial Intelligence / AI and Robotics Research Center

• Collaborated with China Resources Health and Huawei Strategic Research Institute to collect, clean, and preprocess 23,000 surgical case records, building a comprehensive surgical text dataset.
• Expanded Chinese surgical terminology vocabulary, researched data distribution strategies and domain-specific large model training approaches, and participated in incremental training and SFT fine-tuning based on the LLaMA architecture.

Multimodal Reinforcement Learning for Flexible Robotic Bronchoscopy

Institute of Automation, Chinese Academy of Sciences

• Proposed BronchoCopilot, an intelligent agent based on multimodal reinforcement learning for automated robotic bronchoscopy examinations, achieving over 90% success rate in tests of 12 lesion locations across 4 cases, while reducing average contact force by 48.3%.
• Designed multimodal fusion methods and training strategies combining robot vision, pose, and tactile information, training the agent using PPO algorithm. The system has been implemented in practical applications.

RSS Autonomous Driving Safety Framework Development

DiDi / Voyager

• Contributed to autonomous driving safety algorithm research, modeling driving safety scenarios based on the RSS (Responsibility-Sensitive Safety) model, with the technology integrated into Voyager's vehicle algorithm library.
• Trained vehicle agents using imitation learning, enabling the RSS module to constrain and correct driving decisions, resulting in a 79.3% reduction in accident rates in simulator testing.

GNN-based Autonomous Vehicle Trajectory Prediction

Shanghai Artificial Intelligence Laboratory

• Developed a vehicle trajectory prediction model based on Graph Attention Networks (GAT), modeling multi-vehicle motion relationships in traffic scenes.
• Implemented graph-based representation of traffic scenarios to capture vehicle interactions and spatial-temporal dependencies, achieving accurate trajectory forecasting.