MIT-IBM Watson AI Lab

AI models are getting bigger but do they need to be? 🎧 Listen to the full episode of Curiosity Unbounded to learn how efficiency could be the next big breakthrough in AI.

Tess Smidt, an associate professor in @MITEECS, has been named to the 2025 cohort of AI2050 Early Career Fellows. eecs.mit.edu/tess-smidt-nam…

MIT and @MITIBMLab researchers have developed an expressive architecture that provides better state tracking and sequential reasoning in LLMs over long texts. news.mit.edu/2025/new-way-t…

MIT researchers have developed the “self-steering” DisCIPL system, which uses an LLM to plan how to answer complex reasoning tasks, like itinerary planning and budgeting, and then divides the legwork of that strategy among smaller language models. news.mit.edu/2025/enabling-…

A new technique from MIT and @MITIBMLab enables LLMs to dynamically adjust the amount of computation they use for reasoning, based on the difficulty of the question. news.mit.edu/2025/smarter-w…

Charting the future of AI, from safer answers to faster thinking: MIT PhD students in the @MITIBMLab Summer Program are pushing AI tools to be more flexible, efficient, and grounded in truth. news.mit.edu/2025/charting-…

MIT researchers have developed a new method that lets LLMs integrate new knowledge efficiently by generating and applying their own “study notes,” enabling them to adapt over time. news.mit.edu/2025/teaching-…

Creating AI that matters: How the MIT-IBM Watson AI Lab is shaping AI-sociotechnical systems for the future. news.mit.edu/2025/creating-…

A new training method from MIT and @MITIBMLab teaches vision-language generative AI models (VLMs) to localize a personalized object. news.mit.edu/2025/method-te…

MIT and @MITIBMLab researchers have developed a universal guide for estimating the performance of an LLM based on smaller models in the same family, saving time, money, and compute. news.mit.edu/2025/how-build…

CodeSteer guides AI models to switch between text and code to solve tough problems. “A trainer may not be better than the star athlete … but the trainer can still give helpful suggestions,” Yongchao Chen says. “This steering method works for LLMs, too.” news.mit.edu/2025/smart-coa…

Dean Anantha Chandrakasan has been named MIT’s new provost, effective July 1. He has served as the dean of the School of Engineering since 2017 and as MIT’s inaugural chief innovation and strategy officer since 2024. news.mit.edu/2025/anantha-c…

A new AI framework from @MIT and the @MITIBMLab supercharges language models, so they can reason over, interactively develop, and verify valid, complex travel agendas. news.mit.edu/2025/inroads-p…

MIT researchers have unveiled an AI-driven adaptive control system for autonomous drones, significantly cutting trajectory errors in the face of unpredictable winds. news.mit.edu/2025/ai-enable…

A new mathematical approach from @MIT researchers allows animators to simulate rubbery and elastic materials in a way that preserves the physical properties of the material and avoids pitfalls like instability. news.mit.edu/2025/animation…

Check out our new work, CycleReward, below! * Product: a drop-in replacement for things like CLIP; f(img, text) --> score * Idea: cycle-consistent mappings are good (encourages descriptive text, penalizes hallucinations), so we train a reward model to check for this

With demand for cement alternatives rising, an MIT team develops a machine-learning method to hunt for new ingredients across the scientific literature. news.mit.edu/2025/ai-stirs-…

🎉 Our paper “M+: Extending MemoryLLM with Scalable Long-Term Memory” is accepted to ICML 2025! 🔹 Co-trained retriever + latent memory 🔹 Retains info across 160k+ tokens 🔹 Much Lower GPU cost compared to backbone LLM arxiv.org/abs/2502.00592

The recipients of the 2025 Ruth and Joel Spira Awards for Excellence in Teaching are Emilio Baglietto, Kaitlyn Becker, Jelena Notaros, and Jonathan Ragan-Kelley. engineering.mit.edu/connect/news/f…

Humans naturally learn by making connections between sight and sound. A new machine-learning model from MIT, @MITIBMLab and others can match corresponding audio and visual data, which could someday help robots interact in the real world. news.mit.edu/2025/ai-learns…