About
Researcher @ Youtu Lab, Tencent
I am Jian Li (李剑), a researcher at Youtu Lab, Tencent, based in Shanghai, China. My current research centers on large language models (LLMs), multimodal LLMs (MLLMs), agents, and reinforcement learning (RL) — in particular, how models can understand multimodal information more efficiently, retrieve and reason more reliably, and act autonomously in complex tasks.
Before moving into the large-model space, I spent several years in computer vision, with representative work on face detection (DSFD, ASFD), temporal action detection (DBG), and efficient network architectures (EMO).
Research Interests
Experience
- Joined Youtu Lab, Tencent as a Researcher.
- Research intern at Tencent WeChat.
- Our team abccba won 1st place in the 2017 Didi-Udacity Self-Driving Car Challenge.
- Started M.S. studies at PCALab, NJUST (Nanjing University of Science and Technology).
- Research intern at ArcSoft.
- Visiting student at the Institute of Computing, Peking University.
News
- TGPO (efficient policy optimization) accepted to ICML 2026.
- RoleArena (multi-agent role-playing environment) accepted to EMNLP 2026.
- AdaMARP accepted to ACL 2026 Findings; Disco-RAG accepted to ACL 2026.
- Released search-agent works SE-Search and Improving Search Agent with One Line of Code.
- VITA-Audio (efficient speech-language model) accepted to NeurIPS 2025.
- Survey Efficient Multimodal Large Language Models published in Visual Intelligence.
- Multiple papers incl. Unveiling the Ignorance of MLLMs accepted to CVPR 2025.
- MMAD (industrial anomaly detection benchmark for MLLMs) accepted to ICLR 2025.
- Fetch and Forge (dataset condensation for detection) accepted to NeurIPS 2024.
- Released survey A Survey on Benchmarks of Multimodal Large Language Models.
- Efficient backbone EMO (Rethinking Mobile Block) accepted to ICCV 2023.
- Temporal action proposal DBG accepted to AAAI 2020.
- Face detector DSFD accepted to CVPR 2019 (700+ citations).
Current Topics
- Search & RAG agents: self-evolving search agents (SE-Search), retrieval-augmented generation (Disco-RAG), AI Search survey.
- Multimodal understanding & evaluation: MLLM benchmarks (MMAD, MLLM Benchmarks survey), video moment retrieval (LLaVA-MR).
- RL & alignment: efficient policy optimization for LLMs (TGPO), reward modeling, role-playing agents.
- Efficiency & distillation: efficient MLLM survey, dataset distillation, knowledge distillation.
Metrics
As of now: ~2900+ citations, h-index 20, i10-index 28(from Google Scholar; values change over time).
Contact
Feel free to reach out about research collaboration, discussion, or internship opportunities: swordli@tencent.com