CoVeR: Drop 92% of Visual Tokens, Keep 93.5% Performance in Multi-View 3D VLM Reasoning

Scobleizer · x · 2026-09-10

CMU Robotics and Meta propose CoVeR, a deterministic, training-free visual token pruner for multi-view 3D reasoning in 2D VLMs.

The problem: Feeding multiple views to a 2D VLM explodes token counts — LLaVA-OneVision-7B yields 8,748 visual tokens at 12 views, and 31% are spatial duplicates of the same surface seen from another camera. Existing attention/feature-based pruners handle semantic redundancy poorly, keeping near-duplicates of prominent regions while leaving the rest of the scene unrepresented.

Method: CoVeR uses 3D coordinates alone — no attention, no visual features, no retraining. Posed images and depth are back-projected into 3D feature points; coverage initialization removes cross-view duplicates via an adaptively-sized voxel grid, and coverage expansion fills underrepresented regions until the exact token budget is met.

Results:

An interactive demo lets you hover any camera to see exactly which tokens are kept; code is coming soon. Robert Scoble notes this "use less data, get similar results" property matters for robotics.

Related event: CMU and Meta Release CoVeR, Training-Free Visual Token Pruning for 3D Reasoning(2 posts)→

Original post →

More from Embodied

Embodied channel →