Type a few words, get a finished 3D design. Modern AI makes that easy, and it also hides how the result was made. A University of Sydney team says that is a problem for learning, and on November 4, 2026 it announced a tool built to fix it.
The system, called Treellis, was developed in the university’s Faculty of Engineering with the University of Cambridge and Singapore Management University. The paper describing it won a Best Paper Award at ACM UIST 2026, the Symposium on User Interface Software and Technology, held in Michigan.
AI as a black box, and why it matters
The university frames Treellis around a growing worry: that people who lean on AI end up outsourcing their thinking, which can limit their own learning. A prompt-only tool can output a polished model, but the user may have no idea how its parts fit together, and fixing one detail can break another.
Yixiang Chen, the PhD student who led the paper, put it plainly: the tool does not just make things for you, it shows you how things are made. Associate Professor Anusha Withana, who led development, says Treellis lets users take full control of a design and even make mistakes, which he describes as a big part of how people learn.
How Treellis works
Instead of delivering one final design from a handful of prompts, Treellis works step by step with the user during assembly, mixing manual and AI-assisted operations. It makes each part of a model and the spatial relationships between parts visible. Users can select a component, adjust its size or position through a graphical interface, and immediately see the effect on the whole model.
The interface is open source and free to use in a browser at treellis.art. It can also be plugged into other AI platforms through an application programming interface, so it is not tied to one model.
What the study found
The team tested the idea with two groups. Everyone had to design three 3D objects of rising difficulty: a cup, a desk and a computer monitor, each with specific requirements. A baseline group used a standard prompt-based AI tool. The second group used Treellis.
According to the university, Treellis users actively explored how their designs were assembled and tried small, reversible changes to match the target, which helped them pick up 3D design concepts intuitively. Prompt-only users often did not understand how their models were put together, and their attempts to refine one part introduced new problems elsewhere.
Withana links this to two long-standing gaps in human-computer interaction: the gulf of evaluation, which is the difficulty of understanding how a system works, and the gulf of execution, the difficulty of getting it to do what you want. He says the interactive design narrows both and stops users from handing the whole task to the AI.
Recognition and next steps
Best Paper Awards at UIST go to the top 2 percent of papers, chosen from roughly 260 presented this year. Awards chair Anne Roudaut of the University of Bristol said the winners show originality, scientific rigor and potential for lasting impact. The paper is titled “Treellis: Creating In-Situ Parametric Editing Tools to Progressively Design AI-Generated 3D Models.”
The team hopes to take the interface into clinical settings. The university gives examples such as clinicians interactively designing fitted ankle-foot orthotics for children with cerebral palsy, tailor-made wrist splints for arthritis, or personalized ergonomic grips for rehabilitation equipment. These are stated aspirations rather than deployed uses.
What it means
For educators and workplace trainers, Treellis is a concrete example of AI designed to keep the human in the loop. The lesson reaches beyond 3D modelling: tools that expose their intermediate steps may build skills, while tools that only deliver outputs may not.
For professionals with little design experience, it suggests a route to using generative AI without treating it as magic. The caveat is scale. The reported evidence comes from a controlled study with three design tasks, so how well the approach carries over to other fields and longer-term learning remains to be shown.
Bottom line
Treellis does not argue against AI help. It argues for AI that teaches while it assists. With an open-source release and a prestigious award, it gives other builders a template to try.