Frontier3DBench: Benchmarking Coding Agents for 3D Creation across Objects, Scenes, and Interactive Open Worlds

Frontier3DBench. Benchmarking Coding Agents for 3D Creation across Objects, Scenes, and Interactive Open Worlds. By Xinjie Lin and twelve others, from Navers Lab, Einsia AI, and one other institution. Published on the sixth of November.

Coding agents pair visual understanding with program execution, and that opens a new route to automated 3D creation. By turning images and instructions into editable programs, they bring reconstruction, generation and editing into one workflow. The benchmarks for them, though, split evaluation across different task scopes and protocols. As a result it is unclear how capable these agents are at each level of construction. A score on one benchmark for objects says little about scenes, and neither says much about building a world that has to respond when someone acts in it.

Frontier3DBench measures three levels in one place: single objects, scenes, and interactive open worlds. Evaluation looks jointly at appearance, structure and behaviour, so a result has to look right, be built right, and for worlds, act right.

The benchmark holds one hundred and forty-five tasks across nine categories spanning those three levels. Within object and scene tasks, there are two kinds of request: generating from text, and reconstructing from a reference. That pairing lets the benchmark separate making something plausible from making the specific thing that was shown. Eleven model-harness configurations were run on it, so results reflect a model and the harness wrapped around it together, not a model alone.

Strengths differ by level. GPT-6.1-sol leads at the object and scene levels and has the highest overall score, sixty-six point zero two out of one hundred. GPT-6-astra leads interactive world construction. Every configuration scores lower on reference reconstruction than on text-driven generation in object and scene tasks. Producing plausible content is easier than reproducing specified geometry, layouts and visual detail.

Generated worlds average eighty-three point zero eight out of one hundred on local mechanisms. On motion clearance they reach only twenty point eight two, and on stateful workflows thirty point two six. Individual mechanisms often work. Complete activities, which need coordinated actions and persistent state, remain hard. A single working mechanism is one thing. A whole activity that has to run end to end, while the world keeps track of what has already happened, is where the scores collapse.

Two bottlenecks come out of this. One is reference fidelity: matching what was asked for, not just making something plausible. The other is finishing a coordinated activity in a world, rather than wiring up its parts one at a time.