CoVeR: Coverage-Based Token Pruning for Multi-View 3D Reasoning in VLMs

CoVeR. Coverage-Based Token Pruning for Multi-View 3D Reasoning in VLMs. By Nhat-Tan Bui and nine others, from Carnegie Mellon University and Meta Reality Labs. Published on the eighth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Representing a 3D scene as a set of images taken from multiple views lets a 2D vision-language model reason in 3D, reusing the priors it learned in pre- training, and sidestepping how scarce annotated 3D data is. The cost is thousands of redundant visual tokens, and the bill grows with every view added. Pruning those tokens is the obvious answer, but the existing pruners come in two families, and each is limited in this multi-view 3D setting.

Learned importance methods rank tokens by attention or by encoder features. Redundancy here is fundamentally spatial, so these methods keep near-duplicate tokens from a few prominent regions and leave most of the scene unrepresented. Voxelization methods improve spatial coverage, but they cannot enforce an exact token budget. And as observations from different views overlap in 3D, voxelization saturates, capping how many tokens it retains well below the target.

The authors show that spatial coverage is associated with 3D reasoning performance. If coverage is what matters, the selector should aim for coverage directly. CoVeR does that. It is a deterministic, training-free selector that uses only token coordinates, with no learned signals at all. It chooses tokens that, taken together, cover every region of the scene.

That single objective answers the limitations of both families. CoVeR enforces an exact budget for each scene, which voxelization could not. It breaks through the plateau where voxelization saturates. And it avoids the near- duplicate selections that learned importance makes, because a second token from an already covered region adds nothing to coverage. Being training-free and working from coordinates alone, it can be dropped into an existing model as a plug-and-play module.

Extensive experiments show CoVeR outperforming prior state-of-the-art methods on all three 3D reasoning benchmarks, and generalising as a plug-and-play module across four VLMs. With only about eight percent of the visual tokens, it preserves ninety-three point five percent of the performance obtained with every token. That surpasses the previous state of the art by three point nine percentage points on average across the benchmarks.

The choice of signal is the whole story. Learned rankings look at what seems important and end up clustering, and voxel grids look at space but cannot hit a target. Selecting by where tokens are, with coverage as the goal, keeps the scene represented at a fraction of the cost, with nothing to train.