Douwe Kiela

Career

Douwe Kiela's homepage lists a research position at Google DeepMind and an adjunct professorship in Symbolic Systems at Stanford University. Kiela co-founded Contextual AI and served as its CEO; before that Kiela was Head of Research at Hugging Face and, earlier, a Research Scientist at Facebook AI Research.¹

Research

Kiela's research covers machine learning and natural language processing, with a focus on grounded, multi-agent, and retrieval-augmented language understanding and on better tools for evaluation and benchmarking. Kiela co-authored the 2017 InferSent paper on learning general-purpose sentence representations from natural language inference data.¹⁵

Much of Kiela's work concerns how models are tested. Kiela was a co-author of Adversarial NLI (2019), a benchmark collected through an iterative human-and-model-in-the-loop process, first author of the Hateful Memes Challenge (2020) for multimodal classification, and first author of Dynabench (2021), a platform for dynamic dataset creation and benchmarking.⁶⁷³

Kiela was also an author of the May 2020 paper that introduced retrieval-augmented generation, which pairs a pre-trained language model with a retrievable non-parametric memory. Later work includes KTO (2024), an alignment method based on prospect theory, and OLMoE (2024), a family of open mixture-of-experts language models.²⁸⁴

Work and contributions

• OLMoE Open weights

  Co-author of "OLMoE: Open Mixture-of-Experts Language Models" (2024); the abstract says the team open-sourced the model weights, training data, code, and logs.⁴

• Retrieval-Augmented Generation (RAG) Published research

  Co-author of "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (2020), which combined pre-trained language models with an explicit non-parametric memory for knowledge-intensive tasks.²

• Dynabench Published research

  First author of "Dynabench: Rethinking Benchmarking in NLP" (2021), which introduced an open-source platform for human-and-model-in-the-loop dataset creation and model benchmarking.³

Timeline

1. May 5, 2017Co-authored the InferSent paper on universal sentence representations from inference data.⁵
2. Oct 31, 2019Co-authored Adversarial NLI, a benchmark for natural language understanding.⁶
3. May 10, 2020Released the Hateful Memes Challenge paper as first author.⁷
4. May 22, 2020Co-authored the paper introducing retrieval-augmented generation (RAG).²
5. Apr 7, 2021Released Dynabench, Rethinking Benchmarking in NLP, as first author.³
6. Feb 2, 2024Co-authored KTO, a model alignment method framed through prospect theory.⁸
7. Sep 3, 2024Co-authored OLMoE, Open Mixture-of-Experts Language Models.⁴

Other documented roles

Roles the sources document that are not confirmed as current.

• Co-founder and CEO, Contextual AI¹
• Head of Research, Hugging Face¹
• Research Scientist, Facebook AI Research (Meta)¹

Where they work

• Google DeepMind 64 people in the wiki
• Stanford University 40 people in the wiki

Related people

• João CarreiraAlso at Google DeepMind
• Sahil DuaAlso at Google DeepMind
• Dani YogatamaAlso at Google DeepMind
• Nicholas CarliniAlso at Google DeepMind
• Gabe PereyraAlso at Google DeepMind
• Mark ChenLanguage models · Multimodal models
• Nanyun (Violet) PengLanguage models · Multimodal models
• Mira MuratiLanguage models · Multimodal models
• Swabha SwayamdiptaEvaluation and benchmarks · Language models
• Akshita BhagiaLanguage models · Evaluation and benchmarks