About me

I am a fifth 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 pretraining, post-training, and benchmarking multi-modal large language models, specifically video language models. 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. Outside of work, I'm into nature, hiking, biking, going to the gym, running, and reading books. I have a small page where I write my thoughts about the books I read.

News

Publications

View all →

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

NeurIPS 2024 D&B

arXiv / Website / Code
TL;DR ▶

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.

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.

arXiv / Demo / Blog
TL;DR ▶

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.

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

TMLR'24

arXiv
TL;DR ▶

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

EMNLP'23 Findings

arXiv / Code
TL;DR ▶

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.

Attentional Mixtures of Soft Prompt Tuning for Parameter-efficient Multi-task Knowledge Sharing

Akari Asai, Mohammadreza Salehi, Matthew E. Peters, Hannaneh Hajishirzi

EMNLP'22

arXiv / Code
TL;DR ▶

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

CVPR'22

arXiv / Code
TL;DR ▶

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

Amirhossein Kazemnejad, Mohammadreza Salehi, Mahdieh Soleymani Baghshah

ACL'2020

ACL Anthology
TL;DR ▶

Paraphrase generation by retrieving similar paraphrase pairs from a pre-existing corpus and editing them using multi-head attention mechanism.