Research Scientist · ByteDance
Building multimodal AI agents that work at industrial scale.
让多模态大模型与智能体在真实业务中大规模落地
About / 关于我
Qi She (佘琪) is a Research Scientist at ByteDance, where he leads the Applied Algorithms, Foundation, and Strategy Algorithms teams for Business Integrity. He drives the large-scale industrial deployment of Multimodal Large Language Models (MLLMs) and Agentic AI, taking AI agents from research to production at scale.
Qi received his Ph.D. from City University of Hong Kong in 2018, during which he was a visiting Ph.D. student at Princeton University. He then joined Intel Labs as a Research Scientist and received the 2020 Intel Gordy Award, the lab’s highest research honor, named after Gordon Moore. In 2025 he was honored with the CityU Young Alumni Award. His research on generative models and representation learning has been published in leading venues including TPAMI, NeurIPS, ICLR, ICML, CVPR, and ICCV.
佘琪，字节跳动研究科学家，现于字节跳动 Business Integrity 负责应用算法、基建与策略算法团队，主导多模态大模型（MLLM）与智能体（Agentic AI）的大规模工业化落地。2018 年博士毕业于香港城市大学，其间赴普林斯顿大学做访问博士生；此后任英特尔研究院（Intel Labs）研究科学家，曾获英特尔研究院最高研究奖 Gordy Award（2020）与香港城市大学 Young Alumni Award（2025）。研究成果发表于 TPAMI、NeurIPS、ICLR、ICML、CVPR、ICCV 等顶级期刊与会议。
Recent Focus / 工作方向
Focusing on the industrial-scale deployment of Agentic AI within ByteDance’s Business Integrity ecosystem:
- Architecting Agentic Foundations: Developing specialized (M)LLMs optimized for reasoning, planning, and autonomous decision-making in high-stakes security scenarios.
- Scaling Autonomous Agents: Leading the end-to-end integration of Multi-Agent systems to handle massive real-time traffic, ensuring robustness and alignment with evolving safety policies.
- Advancing MLLM-Agent Synergy: Pioneering the fusion of (M)LLMs with agentic workflows to solve complex, fine-grained content understanding challenges at scale.
和优秀人才一起探索并落地多模态大模型，在业务场景做深度适配和极致优化：
- 正式多模态大模型算法工程师 / 应用研究员（SFT、RM、RLHF、Pretrain、Agent 框架与应用）
- 实习多模态大模型、计算机视觉、自然语言处理「算法研究实习生」
News / 动态
2026
- 2026-02 📄 We have 3 papers accepted into CVPR 2026: ThinkGen, CodeDance, and UniEdit-I.
- 2026-01 📄 We have 3 papers accepted into ICLR 2026: Video-KTR, TimeSearch-R, and BranchGRPO.
2025
- 2025-11 🚀 Released MammothModa2, a unified AR-Diffusion framework for multimodal understanding and generation.
- 2025-09 📄 Our work on conditional-diversity token pruning for MLLMs is accepted into NeurIPS 2025.
- 2025-08 🚀 Our team has been advancing Agentic AI foundations at scale, driving the large-scale industrialization of autonomous agents and MLLMs for complex Business Integrity scenarios.
- 2025-06 🎉 Honored to receive the 2025 CityU Young Alumni Award! Grateful to my alma mater, CityU EE, for this recognition of my journey from research to industry.
- 2025-06 📄 Our work on exploiting visual cues for token pruning in VLMs is accepted into ICCV 2025.
2024
- 2024-02 📢 Invited as a reviewer for ICML 2024.
2023
- 2023-12 🛠️ Excited to share that I am now leading the Applied Algorithm & Foundation team at ByteDance Business Integrity.
- 2023-08 📄 Our work "Background-aware Classification Activation Map for Weakly Supervised Object Localization" is accepted into IEEE TPAMI 2023.
- 2023-08 📢 Reviewer service: ICLR 2024, NeurIPS 2023, ICML 2023.
Older News / 更早的动态（2016–2022）
2022
- 2022-11 📢 Reviewer service: CVPR 2023, ICLR 2023.
- 2022-05 📄 Our work "PDO-e3DCNNs" is accepted into ICML 2022.
- 2022-03 📄 We have 3 papers accepted into CVPR 2022.
- 2022-02 📄 "Newton Design: Designing CNNs with the Family of Newton's Methods" is accepted into Science China (Information Sciences).
2021
- 2021-07 📄 We have 3 papers accepted into ICCV 2021.
- 2021-07 📢 Reviewer service: ACM MM 2021 (industrial track), ICLR 2022, NeurIPS 2021.
- 2021-04 📄 "An Efficient and Flexible Spike Train Model via Empirical Bayes" is accepted into IEEE TSP.
- 2021-03 📄 We have 3 papers accepted into CVPR 2021.
2020
- 2020-12 📢 Workshop proposal "Continual Learning in Computer Vision" is accepted by CVPR 2021 (selected from 109 proposals).
- 2020-12 📢 Invited as an expert reviewer for ICML 2021.
- 2020-11 📄 Our survey "Generative Adversarial Networks in Computer Vision: A Survey and Taxonomy" is accepted by ACM Computing Surveys.
- 2020-08 📢 Joined ByteDance AI Lab as a Research Scientist.
- 2020-06 📢 We have successfully organized the 1st "Continual Learning in Computer Vision" workshop at CVPR 2020.
- 2020-06 📄 Our paper "IROS 2019 Lifelong Robotic Vision: Object Recognition Challenge [Competitions]" has been published at IEEE Robotics & Automation Magazine (IEEE RAM). The journal is one of four top-tier journals in robotics field.
- 2020-06 📄 Our paper is accepted as the Oral presentation at the conference of Uncertainty in Artificial Intelligence (UAI).
- 2020-05 📄 Our paper "Synthetic-Neuroscore: Using a neuro-AI interface for evaluating generative adversarial networks" has been accepted by Neurocomputing.
- 2020-01 📄 We have 2 papers regarding lifelong robotic vision accepted by ICRA 2020.
🏆 Awards / 奖项
- 2024 ByteDance Monetization Product and Technology Q4 Inspire Award (团队奖).
- 2023 ByteDance Monetization Product and Technology Q2 Inspire Award (团队奖).
- 2021 3rd Place (3/383) in Google Landmark Recognition 2021, the kaggle competition.
- 2020 Recognized as top 10% high-scoring reviewer at NeurIPS 2020.
- 2020 Annual Intel Labs Gordy Awards (the highest annual research award named after Intel’s co-founder Gordon Earle Moore, 戈登·摩尔奖, famous for 摩尔定律).
- 2018-2020 5 times Intel China Awards (2018 Q4、2019 Q1、2019 Q3、2020 Q1、& 2020 Q2).
- 2016 2nd place in the 10th Global Artificial Intelligence Hackathon by Korea ministry of ICT.
- 2016 Outstanding Academic Performance Award.
📝 Blog / 博文
Key lessons from deploying Multimodal Large Language Model agents at scale in ByteDance's Business Integrity ecosystem — covering architecture choices, reliability challenges, and what it takes to ...
Involution introduces a new atomic operation for deep neural networks that inverts the design principles of standard convolution — spatial-specific and channel-agnostic — achieving better performan...