Two former MIT graduate students and a post‑doc now at IBM trace their career breakthroughs to the MIT‑IBM Computing Research Lab.
Hong’s focus is reinforcement learning. While working with EECS professor Pulkit Agrawal, he improved value‑function learning on the Atari game Montezuma’s Revenge, investigating curiosity‑driven exploration. Providing richer reward feedback boosted performance in robotics, large language models, and scientific simulations. At IBM he now builds an enterprise‑grade agent framework, exploring test‑time training, evolutionary computing, and neuroscience‑inspired online model adaptation.
Ko concentrates on trustworthy AI. Funded by the MIT‑IBM partnership from day one of her PhD, she moved from neural networks to foundation models and joined IBM Research after graduation. Her vLLM Hook plugin accesses internal hidden states to compute safety scores—detecting prompt‑injection and hallucination risks—offering a more cost‑effective alternative to low‑rank adapters.
Arunachalam pursues quantum machine learning, bridging theory and near‑term hardware constraints such as nearest‑neighbor connectivity, noise, and limited observables. Collaborations with Isaac Chuang and Kristan Temme led to two notable papers: one delivering rigorous guarantees for Hamiltonian learning, the other showing quantum kernels can outperform classical kernels under widely‑believed hardness assumptions.
Although they work in distinct domains, all three share a drive to translate principled ideas into practical systems—whether curiosity‑driven agents, trustworthy AI inference hooks, or implementable quantum algorithms—laying the groundwork for real‑world “killer applications.”
Review