Zhang-Wei Hong PhD ‘25 traces his fascination with artificial intelligence back to watching DeepMind play Atari, a spark that ultimately led him from MIT to IBM. Hong, alongside Srinivasan Arunachalam and Irene Ko PhD ‘24, used connections fostered by the MIT-IBM Computing Research Lab to translate theoretical research into practical applications.
“Among all the industrial labs, I think MIT-IBM has better academic collaboration policy and opportunity than the others,” Hong says, highlighting how the lab facilitated a seamless transition from academia to addressing real-world challenges in areas like robotics and large language models. This collaborative pathway is now driving innovation at IBM, with the trio applying their expertise to build more robust and inquisitive AI systems.
MIT-IBM Lab Bridges Academic Theory to Industry Application
Zhang-Wei Hong began working at IBM immediately following the completion of his doctorate at MIT in 2025, a transition directly facilitated by connections forged within the MIT-IBM Computing Research Lab. Ko’s research, initially focused on neural networks and foundation models, benefited from a collaborative direction established with advisors from both MIT and IBM. Luca Daniel, Joseph F. Keithley Professor at MIT, and Pin-Yu Chen, IBM Principal Research Scientist, helped shape her work to maximize its real-world impact, bridging the gap between academic exploration and industry standards.
This alignment of goals, developing AI that is safe, robust, accurate, and fair, was a key factor in her decision to pursue industry application. “That really strikes a balance between pure research and something that’s of industry standard or value,” she explains. The lab’s impact extends beyond simply placing researchers in industry positions; it cultivates a mindset geared towards deployment from the outset.
Irene Ko started to work on trustworthy AI with IBM researchers from day 1 in her PhD, because it was funded by MIT-IBM. She notes that the project represents, as far as we know, the first bridge between the deployment and development in trustworthy AI with the inference engines. This early integration of industry perspectives, she says, is what ultimately drew her to IBM after her PhD, fostered during her five years at MIT.
I started to work [on trustworthy AI] with IBM researchers from day 1 in my PhD, because it was funded by MIT-IBM.
Ko, she says, was particularly advantageous since her goals to develop frontier-safe, robust
Hong’s Reinforcement Learning Advances with Curiosity-Driven Exploration
Zhang-Wei Hong’s doctoral work at MIT culminated in techniques to improve reward feedback for artificial intelligence, directly influencing his current research at IBM. He now applies these methods to areas including robotics and large language models, extending beyond theoretical development to practical application. Hong expresses enthusiasm for curiosity-driven exploration, stating, “I’m very excited about curiosity-driven exploration,” a line of inquiry stemming from his graduate studies.
This approach allows AI agents to proactively seek new data, mirroring human inquisitiveness, and enabling tasks like LLM stress-testing and environmental exploration. Hong is currently developing a framework intended to allow reinforcement learning models to refine their own weights during deployment.
He believes this capability represents an advancement for the field, explaining, “If successful, I think that it would be a very useful system and framework for all of the practitioners in reinforcement learning, because it will be the first framework that enables a model to improve — self-evolve their model weights online at a deployment time.” This work incorporates evolutionary computing to optimize exploration and draws inspiration from neuroscience to improve model performance in real-world scenarios. Ko also initiated trustworthy AI research from day 1 in her PhD, supported by MIT-IBM funding, beginning collaboration with IBM researchers at that time. Ko’s work, alongside Hong’s, demonstrates a successful pathway for translating research into tangible enterprise applications, including chart reading and database querying.
The reason I chose to go into industry after my PhD, and IBM specifically, is that I found great joy in the collaboration during my PhD. That process, those five years, gave me very high rewards in personal fulfillment.
Ko, she says, was particularly advantageous since her goals to develop frontier-safe, robust
Ko’s vLLM Hook Enables Trustworthy AI Deployment in Inference
Irene Ko developed a lightweight vLLM inference engine plugin framework that offers significant cost savings compared to alternative trustworthy AI methods by accessing internal model signals like hidden states and activations. This approach analyzes safety scores, specifically identifying the potential for prompt injection and hallucination within large language model decoding processes, unlike methods relying on low-rank adapters which require additional monitoring and modification steps. This early integration of academic theory with industry application allowed Ko to maintain momentum after completing her PhD, leading her to IBM.
Srinivasan Arunachalam’s work demonstrates a similar pattern of interdisciplinary exploration, finding unexpected mathematical connections in seemingly straightforward problems. “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,” he says. Zhang-Wei Hong anticipates the vLLM Hook will provide a valuable framework for reinforcement learning practitioners, enabling models to self-evolve their weights during deployment, a capability he believes will be important for advancing the field.
Among all the industrial labs, I think MIT-IBM has way better academic collaboration policy and opportunity [than the others].
Hong, an IBM research staff member with the MIT-IBM Computing Research Lab who began his PhD at
Arunachalam’s Quantum Insights Drive Hamiltonian & Kernel Research
He prioritized provability over heuristics, a direction shaped by the collaborative environment that allowed theoretical questions to become concrete research paths. This work culminated in a paper on Hamiltonian learning, establishing rigorous guarantees for understanding the dynamics of quantum systems, and another on quantum kernels, offering theoretical support for the advantages of quantum feature spaces over classical counterparts given certain computational challenges. Arunachalam’s approach reflects a broader pattern of interdisciplinary exploration, where unexpected mathematical insights emerge from seemingly straightforward problems.
“Right off the bat, you don’t see it. He notes a personal benefit from the MIT-IBM collaboration, stating.
If successful, I think that it would be a very useful system and framework for all of the practitioners in reinforcement learning, because it will be the first framework that enables a model to improve – self-evolve their model weights online at a deployment time.
Hong, an IBM research staff member with the MIT-IBM Computing Research Lab who began his PhD at
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
