🔥 News

• 2026.07: I joined SimpleAI as Chief Scientist, leading the pretraining of embodied foundation models.
• 2026.06: I received my Ph.D. degree in Engineering from USTC through the USTC-MSRA joint Ph.D. program.
• 2026.05: 🎉 We are thrilled to announce that EMPG has been accepted by ICML 2026! See you in Korea!
• 2026.02: 🔥 We release a blog about Experience-Driven Learning paradigm and share many interesting findings. 🚧 Please find all details in our notion blog.
• 2025.12: 🎉 We are thrilled to announce that WideSearch has been selected as a key benchmark for evaluating agent capabilities in Seed 1.8. We are also honored to contribute to the Seed 1.8 API as core developers of the new context management feature.
• 2025.09: 🔥 We introduce EMPG: a new framework that solves the credit assignment bottleneck in long-horizon agent training by fixing a fundamental flaw in policy gradients. 🚧 Please find all details in our project page.
• 2025.09: 🎉 We're excited to have contributed to MCP Mark, a solid benchmark for stress-testing comprehensive MCP use. We have open-sourced all details in Github. Welcome to join us!
• 2025.08: 🔥 We introduce WideSearch: a new benchmark to test if AI agents can handle large-scale, repetitive information gathering — the real bottleneck in productivity. 🚧 Please find all details in our project page.
• 2025.06: 🎉 Our paper on VLM robustness, "Safeguarding Vision-Language Models: Mitigating Vulnerabilities to Gaussian Noise in Perturbation-based Attacks," has been accepted by ICCV 2025! See you in Hawaii!
• 2025.05: 🔥 Our latest research, DoctorAgent-RL: A Multi-Agent Collaborative Reinforcement Learning System for Multi-Turn Clinical Dialogue, is now available. This work highlights a crucial principle in Human-Agent Interaction: Agents must proactively request necessary information to excel, as humans may not always volunteer it. This "Agent-must-ask" paradigm is central to DoctorAgent-RL's ability to facilitate better task completion in complex multi-turn dialogues.
• 2025.04: 🎉 Thrilled to kick off my new internship with the ByteDance Seed Team!
• 2025.04: 🔥 Our latest work on VLM robustness, "Safeguarding Vision-Language Models: Mitigating Vulnerabilities to Gaussian Noise in Perturbation-based Attacks," has been released on Arxiv. We've open-sourced the Robust-VLGuard dataset and DiffPure-VLM defense.
• 2025.03: 🎉 Our paper "UniHDSA: A Unified Relation Prediction Approach for Hierarchical Document Structure Analysis" has been accepted by Pattern Recognition Journal!
• 2024.12: 💻 We've launched a new GitHub project: Awesome-Multimodal-RAG! Check out the latest in multimodal RAG and contribute!
• 2024.12: 🤝 We're excited to have contributed to DeepSeek-VL2, an advanced Vision-Language Model with strong performance and fewer parameters.