How the MIT-IBM lab turns quantum and AI theory into real deployed systems

[Millie82]

MIT News profiled three researchers, Srinivasan Arunachalam, Zhang-Wei Hong and Irene Ko, who moved from MIT graduate research through the MIT-IBM Computing Research Lab, formerly the MIT-IBM Watson AI Lab, into research roles at IBM spanning quantum machine learning, reinforcement learning and trustworthy AI. Hong, who began his MIT PhD in 2020, worked on improving reinforcement learning techniques using Atari games as a testbed, and now develops infrastructure for IBM's agentic framework covering tasks like chart reading and database tool calling for enterprise use

Ko's research focuses on trustworthy AI, and she built a tool called vLLM Hook that provides access to a language model's internal signals, like hidden states and activations, to analyze safety scores including the likelihood of prompt injection attacks and hallucination without needing separate low rank adapters layered on top. She described it as the first real bridge between deployment and development specifically for trustworthy AI within inference engines, offering meaningful cost savings compared to other existing monitoring approaches

Arunachalam, who joined MIT as a postdoc in 2018 studying quantum computing theory under physicist Aram Harrow, moved to IBM to work on problems implementable on near term quantum hardware given real constraints like nearest neighbor architecture and hardware noise, contributing to papers on Hamiltonian learning and quantum kernels that gave theoretical evidence for where quantum computing could offer genuine advantages over classical methods. Curious what people think about this kind of structured academic industry pipeline specifically, does it produce meaningfully better research than researchers working in either pure academia or pure industry alone

From MIT to IBM, expediting AI and quantum deploymentnews.mit.edu

[StringTheory83]

Ko's vLLM Hook tool sounds like exactly the kind of unglamorous infrastructure work that quietly makes trustworthy AI actually practical to deploy at scale rather than just staying a nice theoretical framework nobody implements