Jian Li 李剑
Researcher
Youtu Lab, Tencent · Shanghai, China
I am a researcher at Youtu Lab, Tencent. My research focuses on large language models (LLMs), multimodal LLMs (MLLMs), agents, and reinforcement learning (RL). Earlier, I worked extensively on computer vision — face/object detection, temporal action detection, and efficient network architectures. This site collects my research, publications, and technical notes.
- Citations
- 2900+
- h-index
- 20
- i10-index
- 28
Research Interests
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.
Selected Publications
All publications →- SE-Search: Self-Evolving Search Agent via Memory and Dense RewardFirst author
- Agent
- RL
- RAG
- TGPO: Efficient Policy Optimization through Sequence Anchor and Information Gating
- RL
- LLM
- LLaVA-MR: Large Language-and-Vision Assistant for Video Moment RetrievalCorresponding author
- MLLM
- Video
- Efficient Multimodal Large Language Models: A SurveyCo-first author
- MLLM
- Efficient
- Survey
- MMAD: A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly DetectionCorresponding author
- MLLM
- Benchmark
- Anomaly Detection
- A Survey on Benchmarks of Multimodal Large Language ModelsFirst author
- MLLM
- Benchmark
- Survey
- Rethinking Mobile Block for Efficient Attention-based Models (EMO)
- Efficient
- Backbone
- Fast Learning of Temporal Action Proposal via Dense Boundary Generator (DBG)
- Temporal Action Detection
- DSFD: Dual Shot Face DetectorFirst author · Most cited
- Face Detection
Latest Posts
All posts →从策略梯度出发，梳理重要性采样、PPO 裁剪目标，以及 token 级与序列级权重带来的方差差异，作为高效策略优化工作的背景笔记。
以“优化机制”为主轴梳理用 RL 对齐扩散/Flow 生成模型的技术谱系：扩散与 Flow Matching 的数学基础、为何 ODE 要转 SDE、GRPO 家族（Flow-GRPO/DanceGRPO/G²RPO/GRPO-Guard/Flow-DPPO/Flow-OPD）、DPO 与可微反传，以及 Flow-GRPO 与 LLM-GRPO 的本质区别。
我在 ACL 2026 的 oral presentation 整理：面向 RAG 的层次化分块框架 HiChunk 与评测基准 HiCBench。逐页给出幻灯片要点，并附上当时的英文演讲稿（讲稿备注）。
用 verl 真实代码把 6 个主流 RLHF/RLVR 算法一次讲透：以“算法 = Advantage 估计 × Policy Loss”的统一框架，逐一拆解 PPO、GRPO、Dr.GRPO、GSPO、DAPO、OPD，以及 RLOO/ReMax/REINFORCE++/OPO/GPG 等 Critic-free 变体。