Karl Stratos

Karl Stratos is a natural language processing researcher whose work has covered spectral methods, entity linking and pretraining. He completed a computer science PhD at Columbia University in 2016, advised by Michael Collins, and coauthored the 2012 paper Midge: Generating Image Descriptions From Computer Vision Detections, which received an ACL Test-of-Time Paper Award in 2022. After posts at Bloomberg LP, Columbia and the Toyota Technological Institute at Chicago, he was an assistant professor at Rutgers University from 2019 to 2024, where he coauthored EntQA, a model that frames entity linking as question answering. In 2024 he became a member of technical staff at Essential AI, and in 2025 he coauthored its study Practical Efficiency of Muon for Pretraining. He is a researcher on the Apple Foundation Models team.

Now at: Apple
Badges: Bridge

Timeline

2011–2016

Spectral methods and a Columbia PhD

From 2011 to 2016, his research centers on spectral methods, including Brown clustering, word embeddings and sequence labeling. He coauthors Model-based Word Embeddings from Decompositions of Count Matrices with Michael Collins and Daniel Hsu in 2015, and Unsupervised Part-Of-Speech Tagging with Anchor Hidden Markov Models with the same coauthors in 2016. He completes a computer science PhD at Columbia University in 2016, advised by Michael Collins.
2012

The Midge image-description paper

He is one of ten authors of Midge: Generating Image Descriptions From Computer Vision Detections, published at the European Chapter of the ACL conference in Avignon in 2012. His coauthors include Margaret Mitchell, Kota Yamaguchi, Alex Berg, Tamara Berg and Hal Daumé III.
2016–2019

Bloomberg, Columbia and TTIC

After his PhD, he is a senior research scientist at Bloomberg LP from 2016 to 2017 and an adjunct assistant professor at Columbia University in 2017. From 2017 to 2019 he is a research assistant professor at the Toyota Technological Institute at Chicago, where he gives the Research at TTIC seminar The Afterlife of Unsupervised Learning in NLP in October 2017.
2019–2024

Assistant professor at Rutgers

He joins Rutgers University in 2019 as an assistant professor in the Department of Computer Science and stays until 2024, working on natural language processing and machine learning in the Machine Learning research group. His research there centers on entity-centric language understanding, and in 2022 he coauthors EntQA, a model that frames entity linking as question answering, at ICLR with Wenzheng Zhang and Wenyue Hua. With Zhang and Sam Wiseman, he coauthors Seq2seq is All You Need for Coreference Resolution at EMNLP in 2023, which fine-tunes a pretrained seq2seq transformer to output a tagged sequence encoding coreference.
2022

A Test-of-Time award for Midge

He shares the 2022 ACL Test-of-Time Paper Award for 2012 papers with the nine other authors of Midge, including Margaret Mitchell, Tamara Berg and Hal Daumé III.
2024–2025

Essential AI and Muon

He is a member of technical staff at Essential AI from 2024 to 2025. In May 2025, he is among the Essential AI authors, with Ashish Vaswani, of Practical Efficiency of Muon for Pretraining, which shows that the Muon optimizer expands the Pareto frontier over AdamW on the compute-time tradeoff in experiments with models of up to four billion parameters.
As of October 2026

Apple Foundation Models team

He is a researcher on the Apple Foundation Models team, and his research since 2024 centers on pretraining, including optimizers such as Muon and scale invariance.