Shows how language model harnesses can enable compositional generalization across tasks and lengths.
OASYS is a research lab at MIT developing algorithms and abstractions for AI at the last mile. We study how to adapt general-purpose AI models to the downstream context and qualitative feedback that characterize real applications.
We do science in the open. Our work has introduced influential open-source AI models and systems, deployed by many dozens of organizations and collectively downloaded over a hundred million times to date. These include the 
ColBERT retrieval model, the 
DSPy programming framework, and the RLM harness.
Recent Selections
Asks how agents should develop expertise in novel domains from nothing but a corpus.
Introduces a reinforcement learning paradigm that uses privileged information to train models as their own teachers.
Reveals a broad class of natural search tasks that modern retrievers fundamentally fail to tackle.
Introduces an inference-time scaling paradigm that enables language models to process arbitrarily long prompts.
Introduces a prompt optimization algorithm that learns high-level rules from trial and error by reflecting on trajectories.
Introduces an efficient engine for multi-vector retrieval that preserves retrieval quality.
Thoughts on making research impact through open-source models, systems, frameworks, and benchmarks.
People
Omar Khattab
Assistant Professor
MIT EECS & CSAIL
Souradip Chakraborty
Postdoctoral researcher
Alex Zhang
PhD student
Diane Tchuindjo
PhD student
Jacob X Li
PhD student
Matthew Yang
Incoming PhD student
Rikiya Takehi
Incoming PhD student
Noah Ziems
Visiting PhD student
University of Notre Dame
Ziyad Hassan
MEng student
Nathaniel Morgan
Undergraduate researcher
Open Source
Since April 2020
Since Dec 2022