A team of PhD students at MIT working in the MIT‑IBM Watson AI Lab Summer Program are focusing on making AI tools more trustworthy, efficient and grounded in fact. Their projects bridge multiple technical areas — from probing the reliability of large language models to designing new mechanisms for knowledge-grounded reasoning and multimodal inference.
One strand of their work addresses the question of model trustworthiness: for example, student researcher Andrey Bryutkin developed techniques to evaluate when an LLM is likely to give a correct answer — by analysing internal activations like hidden states, token gradients, and sensitivity to out-of-distribution prompts. The idea is to build “meta-probes” that flag potential failures before they’re surfaced to end users.
Another major effort focuses on knowledge-grounded reasoning and retrieval. A second team created a streamlined framework in which a single LLM agent queries a knowledge graph (built from Freebase/Wikidata) via targeted retrieval actions, rather than relying on bulky fixed pipelines. This lets the model reason over structured data more efficiently and with fewer bottlenecks.
Overall, these research efforts reflect a shift in AI from purely bigger models and faster compute toward better behaviour: reducing hallucinations, improving inference efficiency, making outputs more explainable and aligning model actions with real-world knowledge. While the promise of “faster thinking” remains, the emphasis is on anchoring AI’s capabilities in safety, correct reasoning and human-compatible trust.
