Luca Soldaini

Luca Soldaini is a machine-learning researcher whose work spans open language models, multimodal systems and training data. They completed a computer science PhD at Georgetown in 2018, worked on long-tail question answering at Amazon Alexa and coauthored Scim, an interface for skimming scientific papers, in 2022. From 2022 to early 2026, they co-led Ai2's OLMo project; their work included Dolma's three-trillion-token training corpus, Molmo and PixMo, and olmOCR for PDFs. As of October 2026, they are a member of the technical staff at Microsoft AI, working on MAI-Thinking models.

Now at: Microsoft
Links: Website and Google Scholar

Timeline

2018

A Georgetown PhD

They complete a PhD in computer science at Georgetown's Information Retrieval Lab, working with Nazli Goharian in 2018. Their thesis focuses on information retrieval systems for medical experts and lay health users.
2018–2022

Question answering at Alexa

In 2018, they join Amazon Alexa as an applied scientist working on efficient question answering for long-tail knowledge. In 2021, they coauthor an EACL paper on answer selection that reports precision gains of 6% and 11% on two benchmark datasets. They join Ai2 in 2022.
2022

Scim for scientific-paper skimming

In 2022, they coauthor Scim, an interface for skimming scientific papers that highlights salient content to direct readers’ attention.
2022–2026

Co-leading the OLMo project

From 2022 to early 2026, they serve as a lead research scientist at Ai2 and co-lead the OLMo project. During this period, the team releases three generations of dense, mixture-of-experts, hybrid and multimodal model variants, with the data, code, recipes and checkpoints used to build them.
August 2023

Dolma's three-trillion-token corpus

In August 2023, they introduce Dolma, an open corpus of three trillion tokens for language-model pretraining, assembled from web content, academic publications, code, books and encyclopedic material.
2024

Molmo and PixMo

In 2024, they coauthor Molmo and PixMo, presenting an open family of vision-language models and datasets. PixMo includes detailed-caption, image question-answering and pointing datasets collected without external vision-language models.
February 2025

olmOCR for PDF text

In February 2025, they coauthor olmOCR, an open-source toolkit that processes PDFs into clean, linearized plain text while preserving sections, tables, lists and equations. It runs a fine-tuned 7B vision-language model.
As of October 2026

Microsoft AI's MAI-Thinking models

As of October 2026, they are a member of the technical staff at Microsoft AI, working on MAI-Thinking models.