I am a first-year Ph.D. student at the Harbin Institute of Technology, Shenzhen, advised by Prof. Fangyuan Zhang. I received my master’s degree in Electrical and Computer Engineering from the University of Michigan, Ann Arbor, and my B.Eng. in Automation from Zhejiang University.
My research interests include multimodal learning, vision-language understanding, reasoning and generation, and reinforcement learning for foundation models. My recent work focuses on latent visual reasoning in vision-language models and sub-dimensional cross-modal retrieval.
I hold a USPA A License and am a CASI Level 2 snowboard instructor, a PADI Advanced Freediver, and an IKO-certified kitesurfer. I also spend plenty of time with Bajie (八戒), Australian Shepherd, and Golden (狗蛋儿), Maine Coon.
📝 Publications
Decompose, Look, and Reason: Reinforced Latent Reasoning for VLMs
Mengdan Zhu*, Senhao Cheng*, Liang Zhao (*Equal Contribution)
Under Review at ACL ARR 2026
- Premise-conditioned latent reasoning with dynamic multi-step visual grounding
- Spherical Gaussian Latent Policy for RL exploration on hyperspherical manifold
- Three-stage pipeline: contrastive pretraining → SFT → reinforcement finetuning
- V* Bench 83.8%, MathVista 67.5%, MMMU-Pro 56.1%, MMStar 65.2%, surpassing GPT-4o
Cross-modal RAG: Sub-dimensional Text-to-Image Retrieval-Augmented Generation
Mengdan Zhu*, Senhao Cheng*, Guangji Bai, Yifei Zhang, Liang Zhao (*Equal Contribution)
Under Review at ACL ARR 2026 | arXiv | Code
- Sub-dimensional dense retriever with lightweight adaptor (0.01× CLIP’s GPU memory)
- Multi-objective Pareto-optimal image selection with theoretical guarantees
- MS-COCO R@1 81.82% (prev. best 59.10%), Flickr30K R@1 97.50%
ChemSafetyBench: Benchmarking LLM Safety on Chemistry Domain
Haochen Zhao*, Xiangru Tang*, …, Senhao Cheng, …, Mark Gerstein
- Comprehensive benchmark with 30,000+ samples for evaluating LLM safety in chemistry
- Covers chemical properties, usage legality, and synthesis methods
- Incorporates handcrafted templates and advanced jailbreaking scenarios
A Breast Cancer Detection Model Based on Modified ConvNeXt v2
Senhao Cheng, Esther Sun, Wangzi Qian, Yang Han
Published | DOI
- Modified ConvNeXt v2 with Generalized-Mean Pooling and AdaBelief optimizer
- pF1 improvements of 0.031–0.043 over ResNet50, GoogLeNet, and EfficientNet-B2
📖 Education
- 2024 - 2026, M.S. in Electrical and Computer Engineering, University of Michigan, Ann Arbor
- Focus: Multimodal Reasoning, Vision-Language Models, RL for VLMs
- 2020 - 2024, B.Eng. in Automation, Zhejiang University, Hangzhou, China
💼 Internships
- 2023.09 - 2024.04, AI & Data Analysis Intern, MindRank Ltd., Hangzhou, China
- Knowledge Graph construction, Biomedical Data analysis, Drug Discovery pipeline, Predictive Modeling