Wenlong Zhang

Young Researcher of AI for Science Center, Shanghai AI Laboratory
zhangwenlong@pjlab.org.cn
Google scholar | Github | Researchgate
I am a Young Researcher of Shanghai AI Laboratory, working with Prof. Wanli Ouyang and Prof. Lei Bai. Before that, I got the PhD degree from Hong Kong Polytechnic University, working with Prof. Xiao-Ming Wu. I also interned at XPixel group in Shanghai AI Laboratory and SIAT-CAS, working with Prof. Chao Dong and Prof. Yu Qiao. In 2018, I got the Master degree from the Beijing Institute of Technology, supervised by Prof. Weidong Hu.
Currently, I lead a team engaged in scientific discovery evaluation and open-source platforms. We have launched the SciPrismax project. Recently, my primary areas of focus include:
-
Large-scale evaluation
: How do we define, evaluate, and utilize the key capabilities of large models and agents in scientific discovery? -
Post-training algorithm
: How do we enhance multimodal understanding and reasoning with scientific preferences through reinforcement learning and tool call? -
Human-AI collaboration
: How do we align the knowledge of scientists with foundational models and agents? -
Scientific foundation model
: How do we formulate understanding, generation, and reasoning into a unified model for science?
Recent Research Highlights:
-
Large-scale evaluation
:- Multi-modal understanding: Scientists’ First Exam(Preprint), MSEarth(Preprint), OmniEarth-Bench(Preprint)
- Multi-modal reasoning: EarthSE(Preprint), PhysUniBench(Preprint)
- GUI operation: VeriGUI(Preprint)
-
Post-training algorithm
: -
Human-AI collaboration
: -
Scientific foundation model
:- Weather unified model: WeatherGFM(ICLR2025)
- Vision unified model: Lumina-omnilv(Preprint)
- Vision representation: Decouple to Reconstruct(ICCV2025)
If you are interested in the above research topics and would like to join us with Young Researcher, Research Intern or Joint Training Ph.D. Project at Shanghai AI Laboratory, feel free to drop me an email zhangwenlong@pjlab.org.cn. Students with good foundations in AI and science background are appreciated.
news
Jun 12, 2025 | We have released the large-scale multimodal scientific benchmark for scientific discovery scenarios, Scientists’ First Exam. Welcome to download and try it on Hugging Face. One papers were accepted by ICCV. |
---|---|
Nov 08, 2024 | Two papers were accepted by ICLR2025. Our WeatherGFM is the first generalist weather foundation model that can flexibly handle more than 10 weather understanding tasks. It outperforms ECMWF Integrated Forecasting System (IFS) forecast results. |
Nov 07, 2024 | One papers were accepted by ICASSP2025. We propose the first diffusion-based method, DiffSR, to synthesize weather radar data from meteorological satellite data. |
Aug 21, 2024 | One papers were accepted by EMNLP2024. One papers were accepted by NeurIPS. |
Jul 04, 2024 | One papers were accepted by ECCV. One papers were accepted by ACM MM. |
Sep 22, 2023 | One papers were accepted by NeurIPS. One papers were accepted by ICLR and was selected for an Spotlight presentation in December. |
Nov 20, 2022 | One papers were accepted by ICLR. One papers were accepted by CVPR workshop. |
Jul 20, 2021 | One papers were accepted by T-PAMI. One papers were accepted by NeurIPS. |
May 20, 2019 | One papers were accepted by ICCV and was selected for an oral presentation. |
selected publications
- arXivScientists’ First Exam: Probing Cognitive Abilities of MLLM via Perception, Understanding, and ReasoningIn arXiv , 2025
- arXiv
- arXiv
- arXiv
- arXivPhysUniBench: An Undergraduate-Level Physics Reasoning Benchmark for Multimodal ModelsIn arXiv , 2025
- arXiv
- arXiv
- arXivOmniEarth-Bench: Towards Holistic Evaluation of Earth’s Six Spheres and Cross-Spheres Interactions with Multimodal Observational Earth DataIn arXiv , 2025
- arXiv
- arXivAlign-DA: Align Score-based Atmospheric Data Assimilation with Multiple PreferencesIn arXiv , 2025
- arXivEarthSE: A Benchmark Evaluating Earth Scientific Exploration Capability for Large Language ModelsIn arXiv , 2025
- arXiv
- ICCVDecouple to Reconstruct: High Quality UHD Restoration via Active Feature Disentanglement and Reversible FusionIn Accepted to IEEE/CVF International Conference on Computer Vision , 2025
- CEEThe operational medium-range deterministic weather forecasting can be extended beyond a 10-day lead timeIn Communications Earth & Environment , 2025
- ICLRWeatherGFM: Learning A Weather Generalist Foundation Model via In-context LearningIn International Conference on Learning Representations , 2025
- ICLRPostcast: Generalizable postprocessing for precipitation nowcasting via unsupervised blurriness modelingIn International Conference on Learning Representations , 2025
- ICASSPDiffSR: Learning Radar Reflectivity Synthesis via Diffusion Model from Satellite ObservationsIn International Conference on Acoustics, Speech, and Signal Processing , 2025
- CoLM
- arXiv
- EMNLPUniFashion: A Unified Vision-Language Model for Multimodal Fashion Retrieval and GenerationIn Empirical Methods in Natural Language Processing , 2024
- MM
- NeurIPSGeneralizing Weather Forecast to Fine-grained Temporal Scales via Physics-AI Hybrid ModelingIn Advances in Neural Information Processing Systems , 2024