Accelerating AI and Quantum Deployment from MIT to IBM

For readers curious about how academic AI research becomes real products, the path taken by three former MIT researchers now at IBM offers a concrete answer. Srinivasan Arunachalam, Zhang-Wei Hong PhD ’25, and Irene Ko PhD ’24 each spent their formative research years connected to the MIT-IBM Computing Research Lab (formerly the MIT-IBM Watson AI Lab). That connection, they say, helped them close the gap between education and employment and pushed their ideas toward business impact. They now work on quantum machine learning, reinforcement learning and AI agents, and trustworthy and fair AI, respectively.

A lab built to move ideas into industry

The MIT-IBM Computing Research Lab functions as a conduit for research relationship building and for moving expertise into industry applications. Hong, an IBM research staff member with the lab, began his PhD at MIT in 2020 in the Department of Electrical Engineering and Computer Science (EECS). He argues the lab stands out among industrial labs for its academic collaboration policy and opportunity.

That policy matters because the three researchers arrived with very different questions. What unites them is an instinct to move ideas across the space between what is possible in principle and what is useful in practice. Each is chasing what they describe as “killer applications” — real-world use cases that prove underlying research matters beyond the lab.

Reinforcement learning that learns like a person

Hong traces his interest to the moment he learned DeepMind could play Atari and learn from raw screen pixels through feature engineering. During graduate work with EECS Associate Professor Pulkit Agrawal, also a principal investigator with the lab, Hong worked on improving value function learning for reinforcement learning in video games. He used “Montezuma’s Revenge” in Atari to predict and optimize an agent’s policy performance.

With the lab, Hong developed techniques to ground AI for more realistic applications and provide better reward feedback, applying them to robotics, large language models (LLMs), and reinforcement learning for science. He is especially enthusiastic about curiosity-driven exploration, which lets agents be inquisitive about new data the way humans are. That capability supports tasks from generating test cases to stress-test LLMs to exploring new environments.

Now a mentor to his own students, Hong continues open-ended reinforcement learning research. He investigates test-time training for agents and foundation models, and develops infrastructure for IBM’s agentic framework for enterprise tasks such as chart reading and tool calling for database queries. His work includes evolutionary computing to drive better optimization for exploration and leveraging neuroscience to inform deployment-time model improvement. If successful, he says, it would be the first framework enabling a model to self-evolve its weights online at deployment time — a system he believes would be useful to practitioners across reinforcement learning.

Trustworthy AI, from day one to deployment

Ko’s research has been value-driven from both a personal and professional standpoint. She started working on trustworthy AI with IBM researchers from day one of her PhD because it was funded by MIT-IBM. That was advantageous, she says, since her goals of developing frontier-safe, robust, accurate, and fair AI align with those of MIT and IBM, closing the gap between development and real-world deployment. It struck a balance between pure research and something of industry standard or value.

Her MIT-IBM collaboration ran through her advisor in EECS, Joseph F. and Nancy P. Keithley Professor Luca Daniel, and IBM Principal Research Scientist Pin-Yu Chen. That pairing helped define the direction and parameters of her work to maximize impact, first in neural networks and later with foundation models and LLMs. After graduating in 2024, Ko joined IBM Research as a research scientist.

She chose industry, and IBM specifically, because she found great joy in the collaboration during her PhD. Those five years gave her high personal fulfillment, and she wanted to continue the momentum.

Her current project focuses on pain points in trustworthy methods that are not widely deployed in AI inference platforms. Unlike low-rank adapters, which add extra steps to monitor and modify model behavior, her work on vLLM Hook provides a way to access internal model signals, such as hidden states or activations, for decoding LLMs. This vector acts on transformer modules to analyze safety scores, including identifying the likelihood of prompt-injection and hallucination. Ko has developed a lightweight vLLM inference engine plugin framework to program model internals, which could offer significant cost savings over other methods. She calls it, as far as the team knows, the first bridge between deployment and development in trustworthy AI with inference engines.

Quantum insights hiding in plain sight

Arunachalam has always dabbled in quantum research while constantly exploring other areas of theory, hunting for quantum insights and deep math in unexpected lines of inquiry and papers. As he describes it, you often don’t see it at first — the problem looks vanilla — but further investigation reveals interesting math.

That instinct drew him to MIT as a postdoc in 2018 in the group of Professor Aram Harrow in the Department of Physics. Taking a learning theory-first perspective, Arunachalam looked for target algorithms, subroutines, and circuits where quantum speed-ups might be possible. Conversations with Isaac Chuang, the Julius A. Stratton Professor in Electrical Engineering and Physics and an MIT-IBM PI, led him to collaborate with the lab and IBM researcher Kristan Temme.

His transition to IBM was seamless, and he involved himself more closely with problems potentially implementable on near-term quantum devices, keeping constraints in mind: nearest-neighbor architecture, noise, and simpler observable measurements. He focused on quantum machine learning and areas where quantum computing would be superior to classical computing, increasingly prioritizing provability grounded in theory over heuristics. The MIT-IBM connection helped turn theoretical questions into concrete research directions.

That work culminated in two prominent papers. One, on Hamiltonian learning, gave rigorous guarantees for learning the dynamics of quantum systems. The other, on quantum kernels, provided theoretical evidence that quantum feature spaces can offer advantages over classical kernels under widely believed hardness assumptions. Arunachalam also pours himself into different branches of computer science to uncover structure others may miss. Exposing connections between different fields, he says, has let him explore learning quantum states — from completely classically simulatable quantum objects to extremely complicated ones.

What this means for readers

The three careers point to a practical pattern: long-running academic-industry partnerships can shorten the distance between a theoretical result and a deployed system. For students and early-career researchers weighing graduate programs, the lesson is that funding structures and advisor networks shape whether work stays theoretical or finds real constraints. For teams building AI products, the work on vLLM Hook and deployment-time self-evolution suggests where trustworthy and adaptive systems may be heading next. The open question is how quickly these frameworks move from research prototypes into the inference platforms companies already run.