Matt Jordan

Matt Jordan is a machine learning researcher at the Allen Institute for AI (Ai2), where he works on the training data behind its fully open Olmo language models. He earned an undergraduate degree at MIT in 2015 and cofounded MaestroQA, a New York startup for customer support software, before graduate study at UT Austin with Alex Dimakis. There he studied adversarial robustness, presenting the GeoCert certification method in 2019 and an exact algorithm for the local Lipschitz constants of ReLU networks in 2020. In 2024, he coauthored the DataComp-LM benchmark for language-model training data and Ai2's OLMo 2 models. His later work includes data acquisition and curation for Olmo 3 in 2025 and the Olmix data-mixing framework, presented at ICML 2026.

Now at: Allen Institute for AI (Ai2)
Works: 14
Citations: 142
h-index: 5
Badges: Founder
Bridge

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2015

An MIT degree and MaestroQA

He earns his undergraduate degree from MIT in 2015. Before graduate school at the University of Texas at Austin, he cofounds MaestroQA, a New York tech startup that builds customer support management software.
2019

GeoCert for adversarial examples

As a computer science graduate student at UT Austin advised by Alex Dimakis, he researches adversarial examples and spends summer 2019 as a visiting graduate student in the Simons Institute's Foundations of Deep Learning program. With Justin Lewis and Dimakis, he presents GeoCert at NeurIPS 2019, an algorithm that finds the largest norm ball around an input within which a ReLU network's output class stays unchanged.
2020–2022

Lipschitz constants of ReLU networks

In 2020, he and Dimakis show strong inapproximability results for estimating the Lipschitz constants of ReLU networks and give an algorithm that computes them exactly, published at NeurIPS 2020. He follows with Provable Lipschitz Certification for Generative Models at ICML 2021 and Zonotope Domains for Lagrangian Neural Network Verification, with Jonathan Hayase, Dimakis and Sewoong Oh, at NeurIPS 2022.
2024

The DataComp-LM benchmark

He coauthors DataComp-LM (DCLM), a NeurIPS 2024 testbed for training-data experiments built on a standardized 240-trillion-token corpus from Common Crawl. Its DCLM-Baseline dataset trains a 7B-parameter model from scratch to 64% five-shot accuracy on MMLU with 2.6 trillion tokens. In 2024, he also coauthors MINT-1T, a multimodal dataset with one trillion tokens.
December 2024

Base models for OLMo 2

In December 2024, he and the OLMo team at the Allen Institute for AI publish OLMo 2, a family of fully open dense language models released with model weights, training data, code and intermediate checkpoints. He works on base model development, including training and data curation.
2025

OLMoASR speech recognition

In August 2025, he coauthors OLMoASR, open speech recognition models trained on OLMoASR-Mix, one million hours of filtered audio-transcript pairs. Across model scales, OLMoASR performs comparably on average to OpenAI's Whisper on short and long-form benchmarks. His 2025 work also includes Datasets, Documents, and Repetitions, a NeurIPS 2025 study of unequal data quality.
December 2025

Data work on Olmo 3

In December 2025, he coauthors Olmo 3, a family of fully open language models at 7B and 32B parameters released with every stage, checkpoint and data point used to build them. On the project, he works on base model data acquisition, data curation infrastructure and experimentation, and mid- and post-training data curation.
2026

Olmix at ICML 2026

In February 2026, he and colleagues at the Allen Institute for AI release Olmix, a framework for mixing data domains during language-model development, and he contributes to its mixing methodology. Its mixture reuse mechanism matches full recomputation of the mix with 74% less compute, and the paper appears at ICML 2026.