Mojan Javaheripi

Mojan Javaheripi is an AI researcher whose work spans efficient Transformer search and Microsoft’s Phi family of language models. She completed a Ph.D. in Computer Engineering at UC San Diego, received the Qualcomm Innovation Fellowship in 2019, and coauthored LiteTransformerSearch, a training-free, hardware-aware search for Transformer language models, published at NeurIPS in 2022. At Microsoft Research, she coauthored the phi-1 paper and wrote the Phi-2 announcement with Sébastien Bubeck in 2023, then contributed to the Phi-3, Phi-4 and Phi-4-reasoning reports in 2024 and 2025. As of October 2026, she leads the mid-training team at Reflection AI, focused on the science of data and scaling, after serving as a Principal Researcher at Microsoft Research and rotating at OpenAI.

Now at: Reflection AI
Works: 41
Citations: 734
h-index: 13
Badges: Bridge

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2017

A UC San Diego fellowship

She receives the Electrical and Computer Engineering Department Fellowship at UC San Diego in 2017. Her Ph.D. is in Computer Engineering from UC San Diego; her thesis focuses on algorithms for the automated design of efficient and robust deep-learning models in vision and NLP.
2019

A Qualcomm Innovation Fellowship

Her research wins the Qualcomm Innovation Fellowship in 2019.
2022

LiteTransformerSearch at NeurIPS

She is first author of LiteTransformerSearch, a training-free, hardware-aware neural architecture search for Transformer language models, published at NeurIPS in 2022. It runs directly on target devices without GPUs; coauthors include Sébastien Bubeck, Farinaz Koushanfar and Debadeepta Dey.
2023

Phi-1 and Phi-2

In June 2023, she coauthors Textbooks Are All You Need, which introduces phi-1, a 1.3-billion-parameter Transformer model for code. In December 2023, as a Principal Researcher at Microsoft Research, she writes the Phi-2 announcement with Sébastien Bubeck; it presents a 2.7-billion-parameter language model.
2023

A NeurIPS challenge talk

In 2023, she gives a talk on the Phi series at the NeurIPS Large Language Model Efficiency Challenge.
2024

Phi-3 and Phi-4 reports

In April 2024, she coauthors the Phi-3 technical report, which introduces phi-3-mini, a 3.8-billion-parameter model trained on 3.3 trillion tokens. In December 2024, she coauthors Microsoft’s Phi-4 technical report, which presents a 14-billion-parameter model built with a training recipe centered on data quality.
2025

Phi-4 reasoning models

In April 2025, she coauthors the Phi-4-reasoning technical report, which introduces a 14-billion-parameter reasoning model and Phi-4-reasoning-plus, a variant improved through outcome-based reinforcement learning. In September 2025, she presents both models at Microsoft Research Forum as a Member of Technical Staff at Microsoft Research AI Frontiers.
As of October 2026

Mid-training at Reflection AI

As of October 2026, she leads the mid-training team at Reflection AI, focusing on the science of data and scaling. Before this role, she serves as a Principal Researcher at Microsoft Research and rotates at OpenAI.