I am a sixth year PhD student in computer science at
The University of Washington,
where I work under
Hannaneh Hajishirzi and
Ali Farhadi and collaborate closely with Ranjay Krishna.
My research spans multimodal foundation and world models, video understanding, long-context modeling, scalable post-training, and large-scale multimodal data systems.
Most recently, I was a research intern on the AI Frontiers team at Microsoft Research, working on active vision.
Before that, I was a part-time student researcher on the PRIOR team at Ai2, where I led long-context post-training of Molmo2,
and a machine learning research intern at Apple.
Prior to UW, I received my B.Sc. in computer engineering from Sharif University of Technology.
I publish under my full name Mohammadreza, and go by Reza among my friends.
I am on the industry job market!
If you think my work aligns with your interests, please reach out at
mrsalehi@cs.washington.edu.
News
06/2026 Started my research internship with the AI Frontiers team at Microsoft Research, hosted by Neel Joshi and John Langford, working on active vision.
06/2026 Molmo2 was selected as a best paper award finalist at CVPR 2026!
02/2026 Happy to release VidNexus, a versatile video sourcing engine behind nearly a PETABYTE of video data powering Molmo2, VideoNet, and SAGE.
02/2026 Molmo2 and VideoNet (work done by my undergrad Tanush!) accepted to CVPR 2026 as Oral and Highlight.
12/2025 Molmo2 is out! I led long-context post-training for this multimodal video model using context parallelism — our 4B variant outperforms Qwen3-VL 8B on long video benchmarks. See more details about my implementation in the paper and here.
05/2023 Our LLM powered gmail assistant won the first place in fixie.ai hackathon.
01/2023 Started my machine learning research internship with the AIML team at Apple, hosted by Sachin Mehta, working on augmenting CLIP with pseudo-labels from expert models.
Instead of generating coordinates as text, MolmoPoint points by selecting the visual tokens that contain the target concept via special grounding tokens, followed by finer-grained subpatch and location tokens. It achieves state-of-the-art image pointing and improves video pointing and tracking.
Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding
C. Clark*, J. Zhang*, Z. Ma*, J.S. Park*, M. Salehi*, R. Tripathi*, S. Lee*, et al. (*core contributors)
An open family of video-language models with strong video understanding and point-driven grounding. I led long-context post-training, scaling to 512+ frames and 65k+ token contexts with context parallelism; our 4B variant outperforms Qwen3-VL 8B on average across long-video benchmarks.
Domain-specific Action Recognition in the VLM Era
T. Yadav, M. Salehi, J.S. Park, H. Hajishirzi, Y. Choi, A. Farhadi, R. Tripathi, R. Krishna
VideoNet: a benchmark of 1,000+ domain-specific actions across dozens of domains that tests video-language models with few-shot in-context examples. Fine-tuning Molmo2-4B on our ~500k video QA training set surpasses all open-weight 8B models on the benchmark.
One Trajectory, One Token: Grounded Video Tokenization via Panoptic Sub-object Trajectory
C. Zheng, J. Zhang, M. Salehi, Z. Gao, V. Iyengar, N. Kobori, Q. Kong, R. Krishna
TrajViT tokenizes videos into panoptic sub-object trajectories instead of space-time patches, cutting tokens by up to 10x while outperforming ViT3D on retrieval and as a VideoLLM encoder.
REALEDIT: Reddit Edits As a Large-scale Empirical Dataset for Image Transformations
P. Sushko, A. Bharadwaj, Z. Lim, V. Ilin, B. Caffee, D. Chen, M. Salehi, C. Hsieh, R. Krishna
A large-scale image editing dataset of real user requests and human-made edits from Reddit. A model trained on it outperforms competitors by up to 165 Elo points in human judgment.
Molmo and PixMo: Open Weights and Open Data for State-of-the-Art Multimodal Models
M. Deitke, C. Clark, S. Lee, R. Tripathi, Y. Yang, J.S. Park, M. Salehi et al.
We present Molmo, a new family of VLMs that are state-of-the-art in their class of openness. The best-in-class 72B model within the Molmo family not only outperforms others in the class of open weight and data models but also compares favorably against proprietary systems like GPT-4o, Claude 3.5, and Gemini 1.5 on both academic benchmarks and human evaluation.
ActionAtlas: A VideoQA Benchmark for Domain-specialized Action Recognition
Mohammadreza Salehi, Jae Sung Park, Tanush Yadav, Aditya Kusupati, Ranjay Krishna, Yejin Choi, Hanna Hajishirzi, Ali Farhadi
A benchmark for evaluating foundation models on complex actions which requires large enough frame sampling rate to understand fast movements, tracking, and understanding sequence of fine movements.
CLIP meets Model Zoo Experts: Pseudo-Supervision for Visual Enhancement
Mohammadreza Salehi, Mehrdad Farajtabar, Maxwell Horton, Fartash Faghri, Hadi Pouransari, Raviteja Vemulapalli, Oncel Tuzel, Ali Farhadi, Mohammad Rastegari, Sachin Mehta
We showed how one can merge any task-specific expert model from open-source model zoos into foundation models (FMs) such as CLIP. This enhances the visual features of FM for dense prediction and localization tasks without collecting any supervised data.
SHARCS: Efficient Transformers through Routing with Dynamic Width Sub-networks
Mohammadreza Salehi, Sachin Mehta, Aditya Kusupati, Ali Farhadi, Hannaneh Hajishirzi
We introduced a new sample adaptive inference method called SHARCS. It routes samples to different sub-networks with varying widths within any transformer network based on the hardness of input sample.
ATTEMPT: Parameter-efficient Multi-task Tuning via Attentional Mixtures of Soft Prompts
Akari Asai, Mohammadreza Salehi, Matthew E. Peters, Hannaneh Hajishirzi
We introduced a new parameter-efficient fine-tuning method based on prompt tuning. In our method, prompts for some source tasks are learnt and for each sample in a new target task an attentional mixture of source prompts is used as the target prompt.
MERLOT Reserve: Multimodal Neural Script Knowledge through Vision and Language and Sound
Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, Yejin Choi
We introduce MERLOT Reserve, which learns from 20 million YouTube videos through all their modalities (audio, vision, and text). Learning from audio helps broadly -- even on single-image tasks like VCR. Our model learns state-of-the-art representations, that also transfer well to video-based tasks in a zero-shot setting.
Paraphrase Generation by Learning How to Edit from Samples
Paraphrase generation by retrieving similar paraphrase pairs from a pre-existing corpus and editing them using multi-head attention mechanism.
Misc.
Outside of work, I'm into biking, hiking, running, going to the gym, spending time in nature, and reading books.
I have a small books page where I write my thoughts about the books I read.