STRIVE: Structured Spatiotemporal Exploration for Reinforcement Learning in Video Question Answering
Structured visual exploration for reinforcement-learning post-training of large multimodal models on video question answering.
Senior Researcher · Microsoft
Mohsen
Fayyaz
Researcher-builder working on foundation models, post-training, multimodal learning, video understanding, and efficient model design.
Recent research
Selected papers
DocSLM: A Small Vision-Language Model for Long Multimodal Document Understanding
A compact vision-language model for reliable understanding of long multimodal documents under memory constraints.
MemLLM: Finetuning LLMs to Use Explicit Read-Write Memory
Fine-tuning LLMs to use explicit read-write memory for knowledge-intensive tasks.
Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders
Stacked temporal attention inside the vision encoder to improve temporal understanding and reasoning in Video-LLMs.
DeltaLLM: Compress LLMs with Low-Rank Deltas between Shared Weights
A post-training compression method that shares weights across Transformer blocks and models their differences with low-rank deltas.
Adaptive Token Sampling for Efficient Vision Transformers
Adaptive token sampling for reducing redundant computation in Vision Transformers.
6,992citations
23h-index
24i10-index
Community
Workshops & Talks
NeurIPS’25
Invited Panelist · VLM4RWD: Causal and Temporal Reasoning for Video Understanding ↗
Mexico City, Mexico
NeurIPS’25
Mexico City, Mexico
CVPR’25
Nashville, U.S.
CVPR’24
Seattle, U.S.
CVPR’23
Vancouver, Canada
CVPR’22
New Orleans, U.S.
ICCV’21
Virtual
CVPR’21
Virtual
CVPR’20
Seattle, U.S.
ICCV’19
Seoul, Korea
Professional profile
Work experience
Microsoft
Berlin, Germany
Senior Researcher
Jun. 2022 – Present
Microsoft
Reading, UK
Research Scientist Intern
Apr. 2021 – Jun. 2022
Bosch Center for Artificial Intelligence
Renningen, Germany
Research Scientist Intern
Sep. 2020 – Mar. 2021
University of Bonn
Bonn, Germany
Doctoral Researcher · Computer Vision Group
2017 – 2022
Sensifai
Remote
Computer Vision / ML Engineer
2016 – 2017
Academic background
Ph.D.
University of Bonn
Ph.D. in Computer Science · Nov. 2017 – Apr. 2022
Summa Cum Laude — graduated with the highest distinction
Grade: 0.0 / 4.0
German grading scale: 1.0 is the highest regular grade and 4.0 is the minimum grade. A grade of 0.0 denotes the highest doctoral distinction.
Thesis: Holistic Video Understanding: Spatio-Temporal Modeling and Efficiency
Supervisor: Prof. Dr. Jürgen Gall ↗
Best PhD Thesis Award · Faculty of Mathematics and Natural Sciences, University of Bonn · 2025