At MIT, a group of PhD students in the MIT-IBM Watson AI Lab are working on shaping the future of AI around three core pillars: trustworthiness, efficiency, and grounded reasoning. They are not just building more powerful models, but exploring how AI systems can be designed to deliver reliable answers, make quicker decisions, and leverage factual knowledge more intelligently.
One of the major challenges the students are tackling is the uncertainty in AI responses. Andrey Bryutkin, a math graduate student, is probing the internal structure of large language models (LLMs) to understand when they are likely to fail. By using “probes” — small neural networks trained alongside the main model — he examines hidden signals (like activation patterns) to flag unreliable or potentially unsafe outputs.
Another student, Jinyeop Song, is working on improving how LLMs use external knowledge bases without inefficiency. His team built a reinforcement-learning framework that lets an AI agent pull in structured data (from sources like Wikidata) dynamically, reasoning step by step to answer questions more accurately while minimizing computation.
Beyond that, the team is rethinking AI architecture itself. Songlin Yang is developing new transformer-style models that use linear attention instead of the traditional softmax mechanism, which reduces computational load and enables better handling of long or evolving inputs. In parallel, Jovana Kondič and Leonardo Hernandez Cano are pushing the frontier of multimodal AI — designing systems that understand and create visual data, like chart images or textures, by converting visual information into code and back.
