Adina Williams

Career

Adina Williams joined Facebook AI Research as a research scientist in October 2018. Meta's people page notes that Williams earned a PhD in the Department of Linguistics at New York University, with doctoral research on the brain basis of syntactic and semantic processing.¹

Research

Williams works in both directions between linguistics and NLP, according to Meta's profile: bringing findings from linguistics and cognitive science to the training, evaluation and debiasing of NLP systems, and using statistical and corpus tools from NLP to find new quantitative, cross-linguistic facts about human languages.¹

Much of Williams's published work concerns datasets and benchmarks. Williams is first author of the MultiNLI corpus paper, accepted to NAACL 2018, and a co-author of Adversarial NLI (ACL 2020) and of Dynabench (NAACL 2021), an open-source platform for dynamic, human-in-the-loop dataset creation and model benchmarking.³⁴⁸

Williams has also co-authored work on measuring social bias in language models, including the HolisticBias descriptor dataset (EMNLP 2022) and ROBBIE (2023), and on evaluation beyond language, such as the IntPhys 2 benchmark for intuitive physics understanding (2025). Williams is a co-author of Llama 2 and is listed among the authors of the 2026 Muse Spark Safety & Preparedness Report.⁵¹⁶⁷

Work and contributions

• Llama 2 Open weights

  Co-author of Llama 2: Open Foundation and Fine-Tuned Chat Models (2023), which develops and releases pretrained and fine-tuned models from 7 to 70 billion parameters.⁶¹

• Muse Spark Safety & Preparedness Report Published research

  Listed among the authors of the Muse Spark Safety & Preparedness Report (2026), which presents Meta's catastrophic-risk evaluations and broader safety analysis for the Muse Spark model.⁷

• MultiNLI corpus Published research

  First author, with Nikita Nangia and Samuel R. Bowman, of A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference (2017), which introduces the MultiNLI corpus for developing and evaluating sentence-understanding models.³

• Adversarial NLI Open data

  Co-author of Adversarial NLI: A New Benchmark for Natural Language Understanding (2019); the facebookresearch/anli repository is published under the Creative Commons Attribution-NonCommercial 4.0 International license.⁴⁹

• Dynabench Open-source code

  Co-author of Dynabench: Rethinking Benchmarking in NLP (NAACL 2021), which introduces an open-source platform for dynamic dataset creation and model benchmarking.⁸

Timeline

1. Apr 18, 2017Released the MultiNLI corpus paper as first author, later accepted to NAACL 2018.³
2. Oct 2018Joined Facebook AI Research as a research scientist.¹
3. Oct 31, 2019Co-authored Adversarial NLI, a benchmark for natural language understanding published at ACL 2020.⁴
4. Apr 7, 2021Co-authored Dynabench, a platform for dynamic dataset creation and benchmarking, published at NAACL 2021.⁸
5. May 18, 2022Co-authored the HolisticBias descriptor dataset paper, published at EMNLP 2022.⁵
6. Jul 18, 2023Co-authored Llama 2, Open Foundation and Fine-Tuned Chat Models.⁶
7. Jun 11, 2025Co-authored IntPhys 2, a benchmark for intuitive physics understanding in synthetic environments.¹
8. May 14, 2026Listed among the authors of the Muse Spark Safety & Preparedness Report.⁷

Where they work

• Meta 106 people in the wiki

Related people

• Jesse DodgeAlso at Meta
• Emmanuel DupouxAlso at Meta
• Thomas ScialomAlso at Meta
• John X. MorrisAlso at Meta
• Eric Michael SmithAlso at Meta
• Ruchir PuriLanguage models · Datasets and data
• Swabha SwayamdiptaEvaluation and benchmarks · Datasets and data
• Kyle LoDatasets and data · Language models
• Marina DanilevskyEvaluation and benchmarks · Language models
• Pradeep DasigiLanguage models · Evaluation and benchmarks