Wenhu Chen [陈文虎 in Chinese]
Researcher at Meta Superintelligence Labs
Email: hustchenwenhu [at] gmail [dot] com / wenhuchen [at] meta [dot] com
Google Scholar / CV (updated in July 25) / Github / Twitter
Biography
Wenhu Chen is currently a researcher at Meta MSL. He works on multimodal pre-training data and evaluation to support Meta's foundation models. He is also an assistant professor at the University of Waterloo (on Leave). He obtained the Canada CIFAR AI Chair Award in 2022. He worked at Google DeepMind from 2021 to 2025, where he contributed to the Gemini multimodel and evaluation efforts. Before that, he obtained his PhD from the CS department of the University of California, Santa Barbara. His research interest lies in natural language processing, deep learning, and multimodal learning. He aims to design models that handle complex reasoning scenarios, such as math problem-solving and knowledge grounding. He is also interested in building more powerful multimodal models to bridge different modalities. He won the prestigious Golden Jubilee Research Excellence Award at the University of Waterloo in 2025. His won the best paper award at TMLR 2025. He received the Area Chair Award in AACL-IJCNLP 2023, the Best Paper Honorable Mention in WACV 2021, and the UCSB CS Outstanding Dissertation Award in 2021.Research Interest
- Reasoning
- Information Retrieval
- Benchmarks and Evaluation
- Generative AI
Research Highlights
1. Benchmarks
- MMMU: A commonly used multimodal evaluation benchmark
- MMLU-Pro: A commonly used language model reasoning benchmark
- BrowseComp-Plus: A commonly used benchmark for deep research models
- TabFact/HybridQA: Benchmarks for structure knowledge grounding and reasoning
- ClawBench: A commonly used benchmark for everyday online task completion
2. LLM Reasoning and Agents
- Program-of-Thoughts: A prompting strategy to use tools to solve complex reasoning tasks
- MAmmoTH/MAmmoTH2/General Reasoner: Advancing reasoning model to solve complex reasoning tasks
- OpenCoderInterpreter/AceCoder: Advanced coding language models for complex tasks
- Verl-Tool: A holistic training framework for diverse tool use
- OpenResearcher/WebExplorer/DCI: Commonly used baselines for deep research models
3. Multimodal Understanding
- Mantis/MAmmoTH-VL: advanced vision-language models with better reasoning skills
- VL-Rethinker/Pixel-Reasoner: advanced vision-language models with better reasoning skills
- VLM2Vec/VLM2Vec-V2: the framework to enable unified and compositional multimodal information retrieval
4. Multimodal Generation
- Re-Imagen/SuTI/Instruct-Imagen: the most effective and efficient and controllable image generation models
- T2V-Turbo/T2V-Turbo-v2: efficient text-to-video generation models
- MagicBrush/OmniEdit/AnyV2V/UniVideo: powerful image and video editing models
- Tuna/Tuna-2: unified multimodal models for understanding and generation
5. Others
- MERT/ChatMusician/YuE: Foundation models for understanding and composing music
- MAP-Neo: Fully open-source language models with high-quality pre-training datasets
- Gemini-2.5: Contributed to the Gemini-2.5 Model at Google DeepMind.
- MiniMax-M2: Contributed to the MiniMax-M2 Model at MiniMax.
TIGER Lab
I direct the Text and Image GEnerative Research (TIGER) lab. My lab is focused on studying different generative models in different modalities including text, images, videos and music. We are committed to building powerful state-of-the-art models for various domains. Our lab is always looking for talented and self-motivated students.
Awards
- 2025: Outstanding Paper Award at TMLR 2025
- 2025: Math Golden Jubilee Award
- 2024: CVPR Best Paper Finalist
- 2023: AACL-IJCNLP23 Area Chair Award
- 2022: Canada CIFAR AI Chair
- 2021: UCSB CS Outstanding Dissertation Award
- 2021: WACV21 Best Student Paper Honorable Mention
