06/11/2026
Our mission to develop AI-enabled, clinically anchored therapeutic experiences goes hand in hand with our mission to contribute to AI-powered 3D content creation.
This research is in collaboration with the Multimedia Interaction and Communication Lab AAST.
We are pleased to share that our new preprint is now live on arXiv:
📄 3D-CBM: A Framework for Concept-Based Interpretability in Generative 3D Modeling
🔗 https://lnkd.in/duPpcgjt
Modern 3D generative models can produce stunning geometry — but they remain largely opaque. A designer cannot tell a model that a chair leg is too thin to bear load, or that a joint is anatomically incorrect, because there is no shared conceptual vocabulary between human intent and model internals.
This work proposes 3D-CBM: a framework that embeds Concept Bottleneck Models directly into a 3D generative architecture, creating a "glass-box" layer of human-interpretable concepts between raw point cloud input and reconstructed output.
Key contributions:
→ A formal mathematical framework extending CBMs to unstructured 3D data (point clouds & meshes)
→ A three-tier concept taxonomy: geometric primitives → structural components → functional attributes
→ A standardized Human-in-the-Loop intervention protocol for test-time concept overriding
→ A proof-of-concept experiment on PartNet chairs achieving 88.8% concept accuracy and a Chamfer Distance of 0.0115
→ A re-encoding validation methodology that closes the HITL loop and rigorously confirms semantic fidelity — going beyond geometric perturbation metrics alone
The intervention analysis reveals an important asymmetry: concept suppression (e.g., removing legs) is highly effective, while concept imposition (e.g., forcing armrests) is counteracted by the unchanged latent embedding — a finding that motivates intervention-aware training as a concrete direction for future work.
This is, to our knowledge, the first work to formally incorporate CBMs into a 3D generative pipeline, filling a gap explicitly identified across the recent interpretability literature.
This research is sponsored by Yubree Labs - formerly VRapeutic.
