MIT
academicMIT's AI research concentrates in CSAIL under the Schwarzman College of Computing umbrella, with the MIT-IBM Watson AI Lab and the Media Lab alongside. Its signature contribution to the frontier is not from-scratch pretraining but the efficiency layer everyone else trains and serves on: Song Han's HAN Lab produced AWQ (MLSys 2024 Best Paper — still the default 4-bit path in vLLM and TensorRT-LLM), StreamingLLM's attention sinks, the QServe/LServe serving stack, and SVDQuant 4-bit diffusion inference.
A second load-bearing line is linear attention: Yoon Kim's group (with PhD student Songlin Yang) drives DeltaNet, Gated DeltaNet, and Log-Linear Attention (ICLR 2026), maintained in the de-facto reference kernel library flash-linear-attention — the substrate for much of the industry's hybrid-architecture experimentation. Jacob Andreas's LINGO group contributes LM-reasoning work (DisCIPL self-steering models), and Pulkit Agrawal's Improbable AI lab anchors robot learning.
Industry entanglement is structural: Han is also an NVIDIA Distinguished Scientist (Jet-Nemotron ships as NVIDIA), the Nunchaku engine is spinning out, and MIT-IBM work feeds the separately-tracked IBM entry. Liquid AI is an independent spinout, not MIT output.
Your notes
People
- Song Han Google Scholar — Associate Professor, EECS; Director, MIT HAN Lab (formerly NVIDIA Distinguished Scientist (concurrent))
- Yoon Kim Google Scholar — Associate Professor, EECS/CSAIL (formerly IBM Research (2020-21))
- Jacob Andreas Google Scholar — Associate Professor, EECS/CSAIL (LINGO group)
- Pulkit Agrawal Google Scholar — Associate Professor, EECS/CSAIL; Director, Improbable AI Lab; Co-Founder, Eka Robotics
- Antonio Torralba Google Scholar — Professor, EECS/CSAIL (formerly Inaugural head, MIT-IBM Watson AI Lab)