UNSW AI research goes global, with compute support from ResetData
What if AI could keep learning new tasks without forgetting the old ones, and without needing more storage? That's the question behind new UNSW research accepted to the Main Conference of Empirical Methods in Natural Language Processing (EMNLP 2026), one of the world's leading AI conferences.
ResetData supported the UNSW research team with access to high-performance GPU infrastructure, which enabled their experiments.
About the research outcome
Some AI models can understand both video and language. They're powerful, but they find it hard to keep adapting as new tasks and domains come along. Each new task can need a different kind of visual reasoning. Storing a separate set of prompts for every task means that storage quickly builds up over time.
The paper, ‘DynaTokens: Controlling Token Dynamics for Continual Video-Language Understanding’, tests a simple idea. Instead of storing prompts for every task, the model learns to generate the right ones when they're needed. It builds on a large existing AI model that stays unchanged, and only a very small part of the system is trained. Each new task needs just kilobytes of new storage.
In testing, DynaTokens was more accurate and forgot less of what it had already learned. It also performed well on types of video it hadn't seen before.
The work was led by Toan Nguyen, a postdoctoral researcher in Flora Salim's team at the UNSW School of Computer Science and Engineering. He worked with Master's student Yang Liu and collaborator Celso de Melo.
See it in action
Supporting Australian researchers
Australia has the talent to lead in AI research. Work like DynaTokens shows what's possible when Australian researchers have access to cutting-edge AI capabilities.
Universities and research organisations are part of what we call Australia's missing middle. These organisations need advanced AI infrastructure but often lack the capital and scale to build it themselves. Backing these organisations is how Australia turns AI infrastructure into AI innovation.