Omar Khattab

Research

Omar Khattab's research asks how to program AI systems that are partly specified in natural language and whose quality and cost can be tuned with language models. Khattab's group develops models, algorithms, abstractions, and benchmarks for declarative AI programming, for combining and scaling retrieval and reasoning, and for learning from sparse or verbal feedback.¹

Khattab's early work centered on retrieval. ColBERT (SIGIR 2020), written with Matei Zaharia, introduced late interaction for passage search, and was followed by ColBERTv2, the PLAID engine, and the WARP multi-vector retrieval engine, which received a Best Paper Award at SIGIR 2025.⁴⁵¹

Open-source work

Khattab is first author of the Demonstrate-Search-Predict framework (2022) and of DSPy, which compiles declarative language model calls into self-improving pipelines and was presented as a spotlight at ICLR 2024. DSPy is released as an MIT-licensed open-source framework.⁶²³¹

Later open-source projects from Khattab's group include GEPA, a genetic algorithm for improving prompts from reflective feedback, and Recursive Language Models, an inference-time scaling approach developed with Alex Zhang and Tim Kraska.⁷¹

Teaching

At MIT, Khattab co-taught Deep Learning with Sara Beery and Kaiming He in Fall 2025 and Advanced NLP with Jacob Andreas and Chris Tanner in Spring 2026, and introduced a new subject, AI Engineering, for Fall 2026. Khattab leads the MIT OASYS lab, whose former mentees have gone on to doctoral programs at Stanford, UC Berkeley, and elsewhere.¹

Work and contributions

• DSPy Open-source code

  First author of "DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines" (2023), the first entry in the repository's citation section. The stanfordnlp/dspy repository is MIT-licensed and describes DSPy as a framework for programming, rather than prompting, language models, with algorithms that optimize prompts and weights.²³¹

• ColBERT Published research

  Co-author with Matei Zaharia of the ColBERT paper (SIGIR 2020) and co-first author of ColBERTv2 (NAACL 2022); Khattab's site describes ColBERT as introducing the late-interaction approach to search.¹

• GEPA: Reflective Prompt Evolution Published research

  Co-author of the GEPA paper (ICLR 2026, oral), a genetic learning algorithm that improves prompts for language model systems by reflecting on feedback in natural language.⁷¹

Timeline

1. Apr 27, 2020Released the ColBERT paper on late-interaction passage search with Matei Zaharia.⁴
2. Dec 2, 2021Released ColBERTv2 as co-first author with Keshav Santhanam.⁵
3. Dec 28, 2022Released the Demonstrate-Search-Predict framework paper.⁶
4. Oct 5, 2023Released the DSPy paper on compiling declarative language model calls.²
5. 2025Taught Deep Learning at MIT with Sara Beery and Kaiming He in the Fall 2025 term.¹
6. Jul 25, 2025Released the GEPA paper on reflective prompt evolution.⁷

Honors and recognition

• SIGIR Best Paper Award (for WARP) · ACM SIGIR Conference on Research and Development in Information Retrieval, 2025¹

Where they work

• MIT 26 people in the wiki

Related people

• Jonathan FrankleAlso at MIT
• Song HanAlso at MIT
• Saadia GabrielAlso at MIT
• Michael CollinsAlso at MIT
• Karen LivescuAlso at MIT
• Michele CatastaAgents · Language models
• Harrison ChaseAgents · Developer tools
• Binyuan HuiAgents · Language models
• Graham NeubigAgents · Language models
• Piero MolinoLanguage models · Developer tools