M.S. in Computer Science · Georgia Institute of Technology
I'm a master's student at Georgia Tech, where I am fortunate to be working with Christopher MacLellan in the Teachable AI Lab and Vijay Ganesh in the Reasoning and Learning Group.
I study how AI systems can keep learning over time without destroying what they already know, something humans do effortlessly. I work on continual learning and meta-learning, in diffusion models and large language models.
Continual learning. I work on consolidating knowledge into a model as it keeps learning. External stores work well, but I believe some of what a model learns has to live in the model itself.
Meta-learning. Today, every decision in training — what to learn from, what to change, how much — is made for the model. I work towards models that make these decisions themselves, starting with structured autonomy: the model makes some of them, within a structure we design.
The two are closely related. The most natural way to consolidate is through what the model already knows — self-distillation or RL in LLMs, or composing learned knowledge in diffusion models. Such updates diverge less from the model, so they forget less, and they hand more of the learning to the model itself.
I also love to teach, and have been a TA for four semesters. Grading in large courses takes TAs a long time, so I have worked on training models to give students feedback in the meantime.
We show that the empirical Fisher of a diffusion model is rank-1 in low-SNR regimes. Using this, we define a rank-1 EWC penalty and show that, combined with replay, the model forgets less. We then show that this formulation is equivalent to one-step MAML under local approximations, and observe the effect directly: the model re-learns previous tasks faster than other methods.
Constraining how much you update is already a form of learning to learn.
The same intuition as the diffusion work, but at LLM scale the Fisher is too expensive to compute, so SCoL lets the model learn it: meta-RL decides which of its own layers to change when it absorbs a passage. Trained only on short contexts, it beats in-context baselines that see the whole passage on LongBench v2.
Consolidation with autonomy: the model decides where new knowledge goes.
Cobweb builds a concept hierarchy incrementally, one example at a time. As a retrieval index it matches dense encoders and stays robust where kNN collapses; CobwebTM extends it to lifelong hierarchical topic modeling and matches or beats prior incremental methods.
The external-memory side: a retrieval structure that is itself continual.
One language model is trained in several study roles over the same examples — comparing problems, identifying subgoals, explaining. Roles that never produce an answer still improve answer accuracy.
Roles hand more of the training to the model itself; next, the model designs its own roles.
Preprint
From a single out-of-distribution query, recover concept prototypes by score-based mode finding across the noise levels of a pretrained diffusion model, then compose them into a product-of-experts teacher.
Once the algorithm is set, the only open question is what to incorporate.
Given a specific misconception, can a model produce the wrong answer a real student holding it would produce, in a single turn? A counterfactual generation problem.
Human-like errors can benchmark and train models; here, for grading theoretical computer science, where partial credit is hard.
* denotes equal contribution.
All publications · also on Google Scholar
Four semesters of TAing Automata and Complexity shaped how I think about learning in both humans and machines — and led directly to my work on LLMs for education.
Georgia Institute of Technology · Head TA, Spring 2025
Delivered two lectures to 300+ students, led 12 review sessions, and wrote homework and exam problems.
Georgia Institute of Technology
Tutored 150+ hours across AI, algorithms, data structures, and discrete mathematics.
I'm always happy to talk about continual learning, meta-learning, or LLMs for education — and I'm glad to hear from students thinking about research. The fastest way to reach me is email.