Eric Michael Smith

Career

Before joining Facebook AI Research, Smith was a machine learning engineer at Blue Apron, where Smith built and maintained the company's forecast of relative order volume for upcoming product offerings, and was a fellow of Insight Data Science. Meta's people page now lists Smith as a research scientist, with research areas including conversational AI, natural language processing and core machine learning.¹

Research

Smith's early FAIR work was on dialogue and text style. Smith was first author of "Can You Put it All Together", an ACL 2020 paper on evaluating how conversational agents blend skills, and of a paper on controlling style in generated dialogue, and co-authored "Recipes for Building an Open-Domain Chatbot" and BlenderBot 3.⁹¹⁰⁵⁶

From 2021 Smith's work turned to measuring and reducing bias and harm in language models, beginning with a study of how names affect generative dialogue models and continuing with the HolisticBias dataset in 2022. Smith later co-authored safety work on Llama models, including Llama Guard 3 Vision, and alignment research such as "The Alignment Waltz" (ICLR 2026).¹¹²⁴¹²

Llama and Muse models

Smith is a co-author of Llama 2 (2023) and a core contributor to The Llama 3 Herd of Models (2024). In 2026 Smith was listed among the authors of the Muse Spark Safety & Preparedness Report and of the technical report for Muse Glimmer 30B, an on-device agentic model.³⁷⁸¹³

Work and contributions

• Llama 2, Llama 3, and Llama Guard 3 Vision Open weights

  Co-author of Llama 2 (2023), listed among the core contributors of The Llama 3 Herd of Models (2024), and co-author of Llama Guard 3 Vision (2024), a multimodal safeguard model fine-tuned on Llama 3.2-Vision.³⁷⁴¹

• 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.⁸

• HolisticBias Open-source code

  First author of "I'm sorry to hear that": Finding New Biases in Language Models with a Holistic Descriptor Dataset (2022), which introduces HolisticBias; the facebookresearch/ResponsibleNLP license file publishes the HolisticBias code under the MIT License.²¹⁴

Timeline

1. Apr 17, 2020Released "Can You Put it All Together", on blending conversational skills, as first author (ACL 2020).⁹
2. Apr 28, 2020Co-authored "Recipes for Building an Open-Domain Chatbot".⁵
3. Sep 7, 2021Published a study, with Adina Williams, on using names to measure and mitigate bias in dialogue models.¹¹
4. May 18, 2022Released the HolisticBias dataset paper as first author.²
5. Aug 5, 2022Co-authored the BlenderBot 3 paper on a deployed conversational agent.⁶
6. Jul 18, 2023Co-authored "Llama 2 - Open Foundation and Fine-Tuned Chat Models".³
7. Jul 31, 2024Listed among the core contributors of The Llama 3 Herd of Models.⁷
8. Nov 15, 2024Co-authored Llama Guard 3 Vision, a multimodal safeguard model.⁴
9. May 14, 2026Listed among the authors of the Muse Spark Safety & Preparedness Report.⁸
10. Oct 17, 2026Listed among the authors of the Muse Glimmer 30B technical report.¹³

Other documented roles

Roles the sources document that are not confirmed as current.

• Machine learning engineer (before joining FAIR), Blue Apron¹

Where they work

• Meta 106 people in the wiki

Related people

• Shengjia ZhaoAlso at Meta
• Emmanuel DupouxAlso at Meta
• Douwe KielaAlso at Meta
• Jesse DodgeAlso at Meta
• Gabe PereyraAlso at Meta
• Aylin CaliskanSafety and alignment · Evaluation and benchmarks
• Diogo AlmeidaLanguage models · Evaluation and benchmarks
• Saadia GabrielLanguage models · Safety and alignment
• Samuel R. BowmanSafety and alignment · Evaluation and benchmarks
• Collin BurnsSafety and alignment · Language models