Computer-Aided Design (CAD) is essential to modern engineering, but converting existing physical shapes into editable digital models remains a labor-intensive task that demands significant expertise. Traditional AI systems typically generate an entire CAD program in one go, without ever checking the intermediate geometry. In contrast, human engineers construct a part feature by feature, verifying after each operation what still needs to be modeled.
To address this gap, researchers from Hugging Face introduce CADENA (Spanish for "chain"), a model that reconstructs a 3D mesh as a parametric CAD program by growing its sequence of operations one at a time. At each step, CADENA executes the partial program and compares the target geometry with what has been built so far, allowing it to see exactly what remains missing—mirroring the human workflow.
The team also releases CADENA-Bench, a new benchmark designed to evaluate reverse-engineering methods on mechanical parts. It comprises 3,396 real mechanical parts from three industrial corpora, de-duplicated and grouped into six part families. Performance is reported per family, so a method's weaknesses remain visible rather than being averaged away. The benchmark is notably challenging: every learned baseline loses roughly half its score when moving from the DeepCAD dataset to real parts.
CADENA outperforms prior methods on the DeepCAD, Fusion 360, MCB, and CADENA-Bench datasets, achieving a 0.910 voxel IoU on BenchCAD's Vision2Code task. The use of reinforcement learning against executed geometry improves accuracy and reduces the invalid rate to just 0.9%, since operations that fail to build earn no reward.
All resources are publicly available: code at github.com/zhemdi/cadena, model weights at huggingface.co/kulibinai/cadena, and the CADENA-Bench dataset at huggingface.co/datasets/kulibinai/cadena-bench.