←
Qianhui Wu54 posts
Qianhui Wu avatar

Qianhui Wu

@5000hui

Principal Researcher @MSFTResearch. Previously @Tsinghua_Uni.

◷ Joined December 2021▧ 1 Photos and videos
Qianhui Wu avatar
Qianhui Wu@5000hui·Nov 10

🚀 Summer 2027 Research Internship Opportunity! We are looking for passionate research interns to join the Deep Learning Group at Microsoft Research in Redmond for a 12-week internship. Our research interests include: 🤖 Computer-Use Agents: Building general-purpose agents that operate across coding, command-line (CLI), and graphical user interfaces (GUI). 🧩 Environment, Task & Verifier Synthesis: Creating scalable ecosystems for agent training and evaluation. 🧠 Reinforcement Learning: Developing scalable RL methods to advance agent capabilities. 🔄 Recursive Self-Improvement: Exploring how agents can continuously improve their own capabilities and learning processes. Who we're looking for: 🎓 PhD students with strong research backgrounds and a passion for advancing agentic AI. 💡 Experience with LLMs, VLMs, reinforcement learning, or agentic systems is a plus. 📄 A strong research track record, demonstrated through publications or impactful projects, is highly valued. If you're excited about these directions, feel free to reach out with your CV and research interests at qianhuiwu@microsoft.com!

Qianhui Wu avatar
Qianhui Wu@5000hui·Nov 6

Excited to share WebHarbor! ⚓ We introduce 153 executable, Dockerized website environments with 3,518 tasks, designed to support reproducible evaluation and RL training for GUI web agents. These environments capture the complexity of real-world websites, featuring rich visual content, authentic login flows, multi-page navigation, and stateful interactions (e.g., carts, bookings, and form submissions). Beyond individual websites, WebHarbor enables large-scale synthesis of cross-interface environments and longer-horizon tasks, opening up new possibilities for training more capable and generalizable digital agents. Huge thanks to all the amazing contributors who made this possible! 🙌 Check it out! 🚀 🔗 aiming-lab.github.io/WebHarb…

⚓ WebHarbor⚓ Introducing the largest open-source environment resources for GUI Web Agents! 🌐 153 real-world websites (Amazon, GitHub, Booking, YouTube, ...) 🎯 3,518 tasks, including complex multi-step and authenticated workflows 🤖 Built for reproducible evaluation and scalable RL training We started with just 15 environments in May 2026. Today, thanks to 57 contributors from 26 institutions, we've scaled to 153 high-fidelity, image-rich environments—the largest open-source collection of its kind, to our knowledge! 🔥 And they're all executable, offline, and fully open-source! 👀 Explore the environments now: aiming-lab.github.io/WebHarb… Why did we build it, and how? 🧵👇 Threads 1/5

Post attachment
Feng Yao avatar
⟳ Qianhui Wu retweeted
Feng Yao@fengyao1909·Nov 6

4-bit all the way from pre-training to post-training. Pre-training done in just 68 hours. Our model factory keeps shipping. Proud of our team!

Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor. Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon. thinkingmachines.ai/news/ink…

Chaoyun Zhang avatar
⟳ Qianhui Wu retweeted
Chaoyun Zhang@vyokky·Nov 4

Huge thanks to everyone who helped bring this project to life. Try it on your own tasks, share what works, and help us push CUAs further. 🌐 Project: beipo.top/blog/northcua/en/ 🤗 Models: huggingface.co/beipo-ai/Nort… 💻 Code: github.com/beipo-ai/NorthCUA 📄 Report: arxiv.org/pdf/2611.02102v1

Qiushi Sun avatar
⟳ Qianhui Wu retweeted
Qiushi Sun@qiushi_sun·Oct 27

🚀 Excited to introduce OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools! 📱🔧 While top mobile agents rely on proprietary trajectories, open efforts have been bottlenecked by fragile emulators and shallow simulators. OpenMobile-2 bridges this gap with an entirely open-source full stack: environments, data, training recipes, and models. 🌐 Try our interactive playground live in your browser: os-copilot.github.io/OpenMob…

0:54
DailyPapers avatar
⟳ Qianhui Wu retweeted
DailyPapers@HuggingPapers·Oct 22

Microsoft Research's WildAWM A scalable pipeline that synthesizes executable computer-use agent environments and tasks using only open-source models, yielding 341 long-horizon env-task pairs and Qwen agents that generalize to 8 unseen benchmarks.

0:46
Qianhui Wu avatar
Qianhui Wu@5000hui·Oct 21

🚀 How do we make AI-generated code not just functional, but secure? 🎯 SecureVibe explores an important challenge in AI-assisted software development: code that passes functional tests can still contain serious security vulnerabilities. 💡 Key takeaway from SecureVibe: improving secure AI coding requires more than additional coding examples. We need to teach agents to think about security, plan for it, and test for it!

Your coding agent’s code can pass every test and still be exploitable, but users rarely know👾 🤔SFT, RL, or OPD: which teaches security best? 🔥SecureVibe helps vibe coding plan, code, and test for security, +6.9 pts BaxBench, +11.5 unseen CWEs SusVibes, +4.1 SWE-bench Verified

Post attachment
Danqing Wang avatar
⟳ Qianhui Wu retweeted
Danqing Wang@dqwang122·Oct 20

Your coding agent’s code can pass every test and still be exploitable, but users rarely know👾 🤔SFT, RL, or OPD: which teaches security best? 🔥SecureVibe helps vibe coding plan, code, and test for security, +6.9 pts BaxBench, +11.5 unseen CWEs SusVibes, +4.1 SWE-bench Verified

Post attachment
Zhaoyang Wang avatar
⟳ Qianhui Wu retweeted
Zhaoyang Wang@zhaoywang_CS·Oct 20

🔥 Excited to share Agent World Model in the Wild (WildAWM)! 🌍 Fully open-source synthesis 🔓: No proprietary models needed! We synthesized 341 high-quality environment–task pairs, addressing 3 challenges in task synthesis: ✅ Solvability: At least one agent can successfully complete each task. 🔥 Difficulty: Challenging even for frontier models, with GPT-6-Astra achieving only ~72.7 pass@8. 🎯 Verifier reliability: 95% agreement between task verifiers and agent-as-a-judge ensemble. 📈 Effective RL training: Trained exclusively in these synthetic worlds, Qwen3.6-35B-A3B achieves an average +10.58-point improvement across 8 OOD benchmarks!

0:46
Qianhui Wu avatar
Qianhui Wu@5000hui·Oct 20

🚀 Introducing WildAWM: We synthesize realistic, controllable environments and long-horizon tasks for computer-use agents across GUI, MCP, and CLI, with execution-based quality gates. Trained on 341 synthesized environment-task pairs, our agents achieve strong OOD generalization across 8 unseen benchmarks, improving over the baseline by 10 points on average. 📄 Paper: arxiv.org/abs/2610.26828 💻 Code: github.com/MSR-Orchard/WildA… 🌐 Project: msr-orchard.github.io/WildAW… 🤗 Data & models: huggingface.co/collections/C…

0:46
Qianhui Wu avatar
Qianhui Wu@5000hui·Oct 19

🚀 Excited to share Self-HPO: Self-Evolving Coding Agents via Harness-Policy Optimization! 💡A coding agent’s performance depends on both its policy and its harness. Self-HPO jointly improves both through an iterative loop!

🤩 Introducing Self-HPO: let the coding agent improve its own harness, then train a stronger policy model with the optimized harness, and repeat the process for recursive improvement! 💪 Self-evolving: no external model required 🎯 Harness search targets the policy's evolving failure modes 📈 Avg over SWE-Bench Verified/Multilingual/Pro: 57.8% → 64.8% 🧵[1/n]

0:43
Qianhui Wu avatar
Qianhui Wu@5000hui·Oct 14

❓Can agents build the environments they learn from? 🚀Check out our AutoEnvScaling: a step toward self-improving agents that co-evolve as environment builders and task solvers. microsoft.github.io/Orchard-…

Recursive self-improvement needs agents that can build their own training environments. But researchers still part of the loop, making it hard to scale. We introduce AutoEnvScaling, which turns Data Flywheel into a terminal task: a proposer agent builds complete Harbor tasks and sandbox. It builds just like human researchers. This enables self-play RL for terminal agents: one model improves as both builder and solver, without human-in-the-loop.

Post attachment
Xiangzhe (Alex) Xu avatar
⟳ Qianhui Wu retweeted
Xiangzhe (Alex) Xu@XiangzheX·Oct 5

Building reliable agents is hard. Even when agents have access to tools with the same capabilities, how those capabilities are exposed can significantly change agent behavior (e.g., consistency and efficiency). Stop by and discuss with us at #colm26!

Post attachmentPost attachment
Qianhui Wu avatar
Qianhui Wu@5000hui·Sep 25

Check out OpenWebRL 🚀 — A fully open framework for training visual web agents with online multi-turn RL on live websites. 👉 openwebrl.github.io

Excited to introduce OpenWebRL 🚀 A fully open framework for training visual web agents with online multi-turn RL on live websites. Today’s strongest web agents often rely on proprietary infrastructure, large-scale closed data, and costly API-based evaluation. OpenWebRL aims to make this training stack open, reproducible, and cost-friendly. With only 0.4K SFT trajectories and 2.2K online RL tasks, OpenWebRL-4B sets a new open-source SOTA on live-web benchmarks including Online-Mind2Web, DeepShop, and WebVoyager, outperforming much larger open-source agents and even matching proprietary systems. We release the full stack: data, models, training framework, and judge models. Paper: arxiv.org/abs/2606.02031 Project: openwebrl.github.io/ Code: github.com/OpenWebRL/OpenWeb… Models & Data: huggingface.co/OpenWebRL

Post attachmentPost attachment
Qianhui Wu avatar
Qianhui Wu@5000hui·Aug 4

Code and data are now available! 🌟 Check it out: github.com/microsoft/Orchard

Orchard is an open-source framework for the research community to train and evaluate AI agents across task types. It reduces complexity while supporting strong performance from smaller models by enabling researchers to reuse the same infrastructure. msft.it/6019a8fqP

0:20