The gap between theoretical research and production systems closes when scientists move fluidly between both worlds. Three researchers who trained at MIT and now work at IBM - Srinivasan Arunachalam, Zhang-Wei Hong, and Irene Ko - have built careers on that principle, using the MIT-IBM Computing Research Lab (formerly the MIT-IBM Watson AI Lab) as a bridge during their formative years. Their work spans quantum machine learning, reinforcement learning agents, and trustworthy AI, each translating rigorous academic foundations into systems with real-world constraints.
From Atari games to enterprise agents
Zhang-Wei Hong, who earned his PhD from MIT's Department of Electrical Engineering and Computer Science in 2025, began his research captivated by DeepMind's ability to learn from raw screen pixels. During his graduate work with EECS Associate Professor Pulkit Agrawal, a principal investigator with the lab, Hong focused on improving value function learning for reinforcement learning, using the game "Montezuma's Revenge" to predict and optimize agent performance.
Hong said the lab's collaboration policy stood out. "Among all the industrial labs, I think MIT-IBM has way better academic collaboration policy and opportunity [than the others]." His curiosity-driven exploration research now informs his work at IBM, where he develops infrastructure for the company's agentic framework for enterprise tasks like chart reading and database queries. He is investigating test-time training that would let models improve their own weights during deployment - a self-evolving system he believes would be "the first framework that enables a model to improve - self-evolve their model weights online at a deployment time."
Building a bridge for trustworthy AI
Irene Ko, PhD '24, started working on trustworthy AI with IBM researchers from day one of her doctoral program because the work was funded by the MIT-IBM lab. Her collaboration with EECS Professor Luca Daniel and IBM Principal Research Scientist Pin-Yu Chen shaped research that moved from neural networks into foundation models and large language models. "That really strikes a balance between pure research and something that's of industry standard or value," Ko said.
After graduating, Ko joined IBM Research to continue the momentum. Her current project addresses a practical bottleneck: trustworthy methods are not widely deployed in AI inference platforms. While low-rank adapters add extra steps to monitor model behavior, Ko's vLLM Hook framework accesses internal model signals - hidden states and activations - to analyze safety scores, identifying risks like prompt injection and hallucination. "I'm very proud of this project because this is really, as far as we know, the first bridge between the deployment and development in trustworthy AI with the inference engines," she said.
Finding quantum advantages through rigorous theory
Srinivasan Arunachalam arrived at MIT as a postdoc in 2018 in the group of Professor Aram Harrow, bringing a learning-theory perspective to the search for quantum speed-ups. Conversations with Professor Isaac Chuang, a lab principal investigator, led him to collaborate with IBM researcher Kristan Temme. Arunachalam described his approach to uncovering structure in problems others might overlook: "Right off the bat, you don't see it. You think, maybe this is just a vanilla problem, and then once you start investigating it further, you find some really interesting math that comes out of it, which I think is pretty cool."
At IBM, Arunachalam shifted toward problems implementable on near-term quantum devices, accounting for constraints like noise and nearest-neighbor architecture. His work produced two prominent papers: one on Hamiltonian learning that gave rigorous guarantees for learning quantum system dynamics, and another on quantum kernels providing theoretical evidence that quantum feature spaces can outperform classical kernels under widely believed hardness assumptions. He continues to move between branches of computer science, applying research to real systems while maintaining provable theoretical grounding over heuristics.
Why this matters for science and research professionals
The common thread across these three careers is not a single technology but a working method: maintaining rigorous theory while building for actual deployment constraints. For researchers considering industry transitions, the MIT-IBM Computing Research Lab model shows that collaborative environments with clear paths to production can shape research agendas toward problems that matter both intellectually and commercially. The resulting work - whether a lightweight inference plugin, a self-evolving agent framework, or provable quantum advantages - demonstrates that the most useful systems often emerge when theory and application develop side by side, not in sequence.
Your membership also unlocks:
