Principal Researcher @MSFTResearch. Previously @Tsinghua_Uni.
Redmond, WA, USA
Joined December 2021
- Tweets50
- Following149
- Followers774
- Likes147
Qianhui Wu retweeted
🚀 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…
Qianhui Wu retweeted
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.
Qianhui Wu retweeted
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
Qianhui Wu retweeted
🔥 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!
Qianhui Wu retweeted
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!
Qianhui Wu retweeted
We’re calling for more reviews now! Join us to build high quality GUI web agent environments! 🤩
⚓ Introducing WebHarbor: Docking Real Websites for Evolving GUI Agent Environments 🚀
Real websites you can actually RL on: reproducible, login-gated flows, testable, sub-second, byte-identical reset, no reCAPTCHA / geo-blocks / drift. 30+ PRs already.
📢 We need reviewers to push us toward 100+ websites.
Details:
🌐 15 real website environments → fully local Docker mirrors
🔓 Login-gated flows (cart, checkout, bookmarks) are actually testable
⚡ Sub-second byte-identical reset → RL-ready
🎬 643 WebVoyager tasks supported out of the box
🌱 Evolving environments that grow with agent capability
🏠 Project page: aiming-lab.github.io/webharb…
🙌 Contribution guide: aiming-lab.github.io/webharb…
💻 Code repo: github.com/aiming-lab/WebHar…
🤗 HuggingFace Dataset: huggingface.co/datasets/Chil…
📖 Reviewer guide: aiming-lab.github.io/webharb…
🔍 Open PRs: github.com/aiming-lab/WebHar…
@zhaoywang_CS, @5000hui
Qianhui Wu retweeted
Why can (V)LMs agents ace coding and math, yet struggle so badly in more complex environments like computer or phone use? 🤔
We find that one key factor lies in models' ability to understand and *simulate* the environment’s dynamics — and propose **Dyna-Mind** to address this!
🧵[1/n]
Qianhui Wu retweeted
✨ Internship Opportunity @ Google Research ✨
We are seeking a self-motivated student researcher to join our team at Google Research starting around January 2026. 🚀 In this role, you will contribute to research projects advancing agentic LLMs through tool use and RL, with the goal of enabling breakthrough applications.
We are particularly interested in PhD students with a strong background in these areas. If interested, please send a brief self-introduction and your CV to [email protected]. Looking forward to connecting with talented researchers in this exciting space!
Qianhui Wu retweeted
🚀 Excited to announce our 4th Workshop on Computer Vision in the Wild (CVinW) at @CVPR 2025!
🔗 computer-vision-in-the-wild.…
⭐We have invinted a great lineup of speakers: Prof. Kaiming He, Prof. @BoqingGo, Prof. @CordeliaSchmid, Prof. @RanjayKrishna, Prof. @sainingxie, Prof. @YunzhuLiYZ, Prof. @furongh to talk about the exciting researches to bring vision to the wild!
🌎Join top researchers tackling real-world vision challenges — from dynamic environments to embodied agents! See you all at #CVPR2025!
#CVPR2025 #ComputerVision #AI
Qianhui Wu retweeted
Thrilled to share that RegMix has been accepted by #ICLR2025! 🎉 Massive shoutout to the incredible co-authors @xszheng2020 @Muennighoff @GuangtaoZ @LongxuDou @TianyuPang1 Jing Jiang @mavenlin!
🙏 Huge thanks to ICLR reviewers for helping us improve RegMix! 🌟 some key Improvements from v1 :
1️⃣ Expanded experiments to 100 domains, and Regression still works extremely well 📈
2️⃣ Conducted 7B model performance over 100B tokens and RegMix beats Human consistently🚀
3️⃣ More results on the confirmation of the rank invariance hypothesis across model scales 📉
4️⃣ New insights at using 1B proxy model level, and it does not show significant advantage than 1M proxy models actually🧠
Code: github.com/sail-sg/regmix
Paper: arxiv.org/abs/2407.01492
Still following your human intuition to mix corpora from different sources for language model pre-training 🧠? Everyone says that data mixture has a big impact on model performance, but how - and why🕵️? Did you know that web corpora are actually highly impactful for downstream tasks 🏆?
Let's check out our preprint "RegMix: Data Mixture as Regression for Language Model Pre-training" 📄
🔬In this paper, we've proposed an automatic data mixture method RegMix that achieves a 6.3% improvement over human selection on the widely used HellaSwag benchmark - and it only needs a 2% extra training FLOPs! 📈
Details in the thread 🧵
Qianhui Wu retweeted
SCBench has been accepted by #ICLR2025! Now you can evaluate your long-context methods across the full KV cache lifecycle. Congratulations to @liyucheng_2 and all my co-authors! Find more details at aka.ms/SCBench
🕸️ KV cache has its own lifecycle, but previous benchs focus solely on single-req, ignore full lifecycle.
🪻SCBench fills the gap with a KV-cache-centric angle, covering 4 kv cache stage across 12 tasks, 4 capability, and 2 shared-context mode
aka.ms/SCBench
Qianhui Wu retweeted
Google presents:
Scaling Embedding Layers in Language Models
Outperforms a 1.9B parameter baseline across diverse corpora, while using only half the inferencetime FLOPS
Qianhui Wu retweeted
Read this blog to learn how dynamic sparse attention in MInference 1.0 enables a single GPU processes 1M context 10x faster.
huggingface.co/blog/liyuchen…
#LLMs #mlsys #TTFT
MInference by Microsoft is released
Milliontokens Inference achieves a 10x speedup for pre-filling and maintains accuracy with 1M tokens ⚡️
GitHub: github.com/microsoft/MInfere…
Discuss with the authors at huggingface.co/papers/2407.0…
Demo: huggingface.co/spaces/micros…
Kudos @iofu728!
Qianhui Wu retweeted
Thanks @_akhaliq for sponsoring. You can now use MInference online in HF Demo huggingface.co/spaces/micros… with ZeroGPU.
Now, you can process 1M context 10x faster in a single A100 using Long-context LLMs like LLaMA3-1M, GLM4-1M, with even better acc, try MInference 1.0 right now!
Qianhui Wu retweeted
🤳I will present LLMLingua at Session 5, Dec. 9 10:15-, and LongLLMLingua at BoF-6, Dec.9 11:00-12:30 #EMNLP2023
llmlingua.com/
Looking forward to meeting you all and talking about prompt compression, efficent method of LLMs, Speculative decoding, etc.
@emnlpmeeting
Advanced prompting technologies for LLMs can lead to excessively long prompts, causing issues. Learn how LLMLingua compresses prompts up to 20x, maintaining quality, reducing latency, and supporting improved UX: msft.it/6017i31xF
