AI News Research

MIT professor Jacob Andreas works on AI models that explain what they learned

MIT News profiles Jacob Andreas, a newly tenured associate professor who develops AI models that can explain what they learned in training.

Automated summary. Source linked.

On November 17, 2026, MIT News published a profile of Jacob Andreas, a newly tenured associate professor in the Department of Electrical Engineering and Computer Science and a member of CSAIL. His research studies the computational basis of how machines learn language. He also develops techniques that help AI models describe accurately what they learned in training.

His group works on two main areas. One studies the science of large language models, including how they learn and process information. Another designs new training algorithms and methods for predicting how specific datasets will affect model behavior. Andreas says the field seems to change every three to five years. He co-teaches courses 6.461 and 6.861, which he redesigns each year.

Andreas says it is hard to predict where the field goes next, with so much change. He expects his group's interpretability work to matter more for making advanced systems such as AI agents safe enough to deploy in high-stakes situations. Outside the lab, he hikes and backpacks, and he is working toward the 4,000-foot peaks of the White Mountains of New Hampshire with a colleague.

Key facts

  • Jacob Andreas recently earned tenure as an associate professor at MIT.
  • Andreas's group develops techniques that let AI models explain accurately what they learned.
  • Andreas says the AI field seems to change every three to five years.
  • Andreas co-teaches courses 6.461 and 6.861, undergraduate and graduate Natural Language Processing.
  • Andreas plans to return to singing in a chamber choir now that he has tenure.

Why it matters

Its group's tools aim to help scientists and researchers anticipate how a trained model will act in the real world. One family of its techniques trains large language models to explain other neural networks in language a human could understand. Andreas expects this interpretability work to grow in importance for advanced systems such as AI agents.

Source: Read it at news.mit.edu (opens in a new tab)

More in this topic