Kyle Lo

Research

Lo's earlier work focused on scientific text and reading. Lo co-authored SciBERT (2019), a language model pretrained on scientific papers, and was first author of S2ORC, the Semantic Scholar Open Research Corpus. Lo also worked on specializing language models to domains, human-AI interaction for sensemaking over large document collections, and augmented reading interfaces, including the Semantic Reader project summarized in Communications of the ACM in 2024.⁶⁷¹

Lo's current research centers on large-scale pretraining, with emphasis on data curation, data mixing, benchmark validity, efficient experimentation and generalization across model scales. Recent papers listed on Lo's site include Signal and Noise, on why some benchmarks make task scaling laws hard to fit, FlexOlmo, a mixture-of-experts model whose experts are trained on separate data sources, and Olmix, the data mixing framework used for OLMo 3.¹

Open models and data

Lo co-led Ai2's OLMo project on fully open language models. Lo was the last-listed author of the Dolma pretraining corpus paper and a co-author of the original OLMo paper, both released in early 2024 and both given best paper awards at ACL 2024. Lo then contributed to OLMoE, the family's first mixture-of-experts model, to OLMo 2 and its paper "2 OLMo 2 Furious", and to OLMo 3, released in November 2025 with its training datasets.¹²³⁵⁴

Lo has also co-authored data tools that support this work, such as olmOCR (2025), which uses vision language models to turn PDFs into text for training corpora. In December 2024 Lo co-presented a NeurIPS tutorial, "Opening the Language Model Pipeline", with Akshita Bhagia and Nathan Lambert. Lo's site marks a farewell to Ai2 in March 2026.⁸¹

Work and contributions

• OLMo Open weights

  Co-led the OLMo project at Ai2, per Lo's site; co-author of the OLMo paper (2024), credited in OLMo 2 with base-model development, and in OLMoE with helping create the pretraining dataset and analyzing data experiments.¹³⁴⁵

• Dolma Open data

  Last-listed author of the Dolma paper, which describes a three-trillion-token English corpus released for research on language-model pretraining.²

• Olmo 3 7B (base) Open weights

  Credited in the Olmo 3 technical report, which the model card links as its paper, for technical leadership, model architecture and training methodology, and base-model data acquisition and curation.⁹

Timeline

1. Mar 2019Co-authored SciBERT, a pretrained language model for scientific text.⁶
2. Nov 2019Posted S2ORC, the Semantic Scholar Open Research Corpus, as first author.⁷
3. Jan 2024Posted the Dolma corpus paper as last-listed author.²
4. Feb 2024Co-authored "OLMo: Accelerating the Science of Language Models".³
5. Aug 2024The OLMo and Dolma papers received best paper awards at ACL 2024.¹
6. Sep 2024Co-authored OLMoE, the first mixture-of-experts model in the OLMo family.⁵¹
7. Dec 2024Co-presented the NeurIPS 2024 tutorial "Opening the Language Model Pipeline" and co-authored "2 OLMo 2 Furious".¹⁴
8. Feb 2025Co-authored olmOCR, on converting PDFs to text with vision language models.⁸
9. Nov 20, 2025Announced on Lo's site the release of OLMo 3 by the Ai2 team, along with its training datasets.¹
10. Mar 26, 2026Announced on Lo's site a farewell to Ai2.¹

Documented roles

The sources do not confirm a current role; these are the roles they document.

• Research scientist; co-lead of the OLMo project, Ai2 (Allen Institute for AI)¹¹⁰

Honors and recognition

• ACL 2024 Best Paper Award (OLMo; Dolma) · Association for Computational Linguistics, 2024¹

Companies in their history

• Ai2 23 people in the wiki

Related people

• Jesse DodgeConnected to Ai2
• Swabha SwayamdiptaConnected to Ai2
• Pradeep DasigiConnected to Ai2
• Mohit IyyerConnected to Ai2
• Sameer SinghConnected to Ai2
• Marina DanilevskyEvaluation and benchmarks · Language models
• Yacine JerniteDatasets and data · Language models
• Rohan TaoriLanguage models · Evaluation and benchmarks
• Eugene YanLanguage models · Evaluation and benchmarks
• Ruchir PuriLanguage models · Datasets and data