Like many AI researchers, I've been intrigued by recent advances in large language models (LLMs), and generative AI more broadly. At the same time – and also in common with many researchers – I'm concerned about their reliability, robustness, and ease of control. This cluster of projects looks into what we can do about that.
The strategy I've been pursuing, with number of collaborators, is to put the LLM inside a larger AI system. Something I picked up from games that I use in thinking about AI is: what's the "core AI loop"? In reinforcement learning it's something like act-observe-update. In chatbot-style AI it's prompt-generate-prompt-generate.
What's a good core AI loop? In my opinion, a lot of "agentic AI" work as of 2025 is too ad-hoc on that question, essentially wrapping an LLM in a hand-crafted loop. I think we can do better by starting with a classical AI algorithm as the core AI loop, then looking for the brittle parts that can be LLM-ified.
Separately, I've grown concerned about reproducibility of published results that use generative AI systems, especially those that use closed-weight, gated models like ChatGPT, so have been doing some work on that too.
Publications:
Funding provided by:
Collaborators (current): Adam Gaier, Amy K. Hoover, Ioannis Koutis, Joel Lehman, Elliot Meyerson, Arash Moradi Karkaj, Ben Samuel, Mike Treanor