body{background:#f5f1e8;color:#242923;font:19px/1.6 Georgia,serif}#root{display:none}.site-static{max-width:940px;margin:auto;padding:32px 24px}.site-static nav{display:flex;justify-content:space-between;gap:24px;padding-bottom:24px;border-bottom:1px solid #cfcec1}.site-static nav span{display:flex;flex-wrap:wrap;gap:18px}.site-static h1{font-size:clamp(42px,7vw,80px);font-weight:400;line-height:1.1;margin:60px 0 24px}.site-static h2{font-size:34px;font-weight:400;margin:40px 0 20px}.site-static h3{font-size:24px;font-weight:400;margin:12px 0}.site-static .lead{font-size:28px}.site-static article{border-top:1px solid #cfcec1;padding:20px 0}.site-static p{margin:12px 0}.site-static a{text-decoration:underline;text-underline-offset:3px}.site-static .meta,.site-static time{color:#60675b;font:16px/1.5 system-ui,sans-serif}.site-static footer{border-top:1px solid #cfcec1;margin-top:56px;padding-top:24px}

Zekun Wang

How can models keep learning after deployment?

I'm a third-year PhD student in Computer Science at Georgia Tech, advised by Christopher MacLellan. I'm also a Research Intern at Amazon in Washington, working on deployment-time agent adaptation and self-improving harnesses.

I study continual learning for deployment-time adaptation and generalization.

zekun@gatech.edu · Google Scholar · About me

News

2026-10 New blog: Continual Learning Diffusion Models. Learning new tasks while preserving earlier abilities.

2026-09 New blog: Self-Consolidating Language Models. Writing context into model weights at test time.

2026-09 Trust Region Continual Learning as an Implicit Meta-Learner was accepted to NeurIPS 2026.

2026-09 I'm co-organizing TTCL, a NeurIPS 2026 workshop on test-time continual learning agents.

2026-08 I joined Amazon as a Research Intern in Washington.

Inside the research

Continual Learning Diffusion Models

Learning new tasks while preserving earlier abilities.

Self-Consolidating Language Models

Writing context into model weights at test time.

Selected publications

arXiv 2026 · 2026

Self-Consolidating Language Models

Zekun Wang, Anant Gupta, Zihan Dong, Christopher J. MacLellan

SCoL post-trains an LLM via meta-RL to emit textual instructions selecting which Transformer layers to LoRA-update from streaming context, balancing acquisition against forgetting.

NeurIPS 2026 · 2026

Trust Region Continual Learning as an Implicit Meta-Learner

Zekun Wang, Anant Gupta, Christopher J. MacLellan

Combines generative replay with a Fisher-metric trust region for continual diffusion training, and shows the update reduces to a one-step MAML form.

arXiv 2026 · 2026

Test-Time Compositional Generalization in Diffusion Models via Concept Discovery

Zekun Wang, Anant Gupta, Tianyi Zhu, Christopher J. MacLellan

Recovers query-specific concepts from a pretrained diffusion model's time-indexed scores via gradient ascent, then composes them as a product-of-experts teacher for test-time generalization.

ICLR 2026 · 2026

Avoid Catastrophic Forgetting with Rank-1 Fisher from Diffusion Models

Zekun Wang, Anant Gupta, Zihan Dong, Christopher J. MacLellan

Proposes a rank-1 Fisher approximation for EWC that exploits the collinear per-sample gradients of diffusion models at low SNR, paired with generative replay for continual image generation.

NAACL 2025 · 2025

Babysit A Language Model From Scratch: Interactive Language Learning by Trials and Demonstrations

Ziqiao Ma, Zekun Wang, Joyce Chai

Trains GPT-2 from scratch with a teacher model giving demonstrations on student trials, scored by an age-conditioned reward, to accelerate word acquisition.

All publications