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

Coverage, not concentration

Rendering a 3D scene as multi-view images lets a pre-trained 2D VLM reason in 3D by reusing its visual and language priors — sidestepping the scarcity of annotated 3D data. But every added view adds visual tokens: 12 views already yield 8,748 tokens in LLaVA-OneVision-7B, and roughly 31% of them are spatial duplicates of the same physical surface seen from a different camera. Existing token pruners weren't designed for this redundancy, which is geometric, not semantic, so ranking tokens by attention or encoder features keeps near-duplicate observations of a few prominent regions and leaves the rest of the scene unrepresented.

Figure 1 — CoVeR. A deterministic, training-free token selector for multi-view 3D reasoning in 2D VLMs. Posed images and depth are back-projected into 3D feature points; coverage initialization removes cross-view duplicates with an adaptively-sized voxel grid, and coverage expansion extends the cover into still-underrepresented regions until an exact token budget is met. Using geometry alone, CoVeR surpasses SOTA token pruners in the multi-view 3D setting.

See what CoVeR keeps (Hover over any Camera)

Results on 3D reasoning benchmarks

We evaluate all three forms of 3D reasoning with LLaVA-OneVision-7B on 12 uniformly sampled views: ScanQA (spatial scene understanding), SQA3D (situated reasoning), and OpenEQA (embodied QA). CoVeR achieves the best aggregate performance at every token budget, and the gap to prior SOTA pruners widens as the budget tightens.

EM@1 / CIDEr / ROUGE-L on ScanQA, LLM-Match on OpenEQA, EM@1 on SQA3D; Avg./Rel. computed across the three benchmarks.

Scaling with views. As more input views are added, CoVeR's ScanQA EM@1 keeps climbing and its lead widens — from +0.7 at 8 views to +2.4 at 24 — while attention- and feature-based pruners flatten out. More views mean more geometric redundancy, exactly what coverage-based selection exploits.

How CoVeR compares

Prior pruners fall into two families. Voxelization methods cover space but can't hit an exact per-scene budget; learned-importance methods lean on attention or encoder features and spend the budget on near-duplicates. CoVeR is the only approach that is geometry-only and guarantees an exact budget on every scene.

Only CoVeR satisfies every property: purely geometric selection with a guaranteed exact token budget on every scene.

Two geometry-only stages

CoVeR selects a token subset C of exact budget B that minimizes the directed Hausdorff distance from every token location X to the retained set. A small value ensures that no region of the scene is dropped outright, which is precisely the guarantee learned importance never offers. CoVeR reaches this objective in two complementary, training-free stages using only backprojected 3D token coordinates — no attention, features, or auxiliary encoder.

# Input:  token coords X = {t_i}, budget B, split ratio α
# Output: selected set C, |C| = B

B_init ← max(1, ⌊α·B⌋)

# Stage 1 — Coverage Initialization
v_s    ← heuristic interval search(lower_bound, higher_bound) s.t. |voxels(X, v_s)| ≈ B_init
C_init ← nearest-to-centroid token per occupied voxel

# Stage 2 — Coverage Expansion
C ← C_init
repeat B − |C_init| times:
    m* ← argmax_{m∉C}  min_{s∈C} ‖t_m − t_s‖²
    C  ← C ∪ {m*}                           # add the least-covered token

return C                                     # |C| = B, exactly

Plug-and-play across VLMs

Because CoVeR needs no training and no architectural changes, the same geometry-only selector drops into different backbones unchanged. We test it across four VLMs — LLaVA-OneVision-7B, Video-3D LLM, Qwen2.5-VL-7B, and Qwen3-VL-8B.

Video-3D LLM backbone, 16 views at a 10% token budget. Geo3DPruner adds a VGGT-1B encoder and retrains the whole backbone; CoVeR adds neither and still comes out ahead on retained performance.

Why coverage wins

Qualitative comparison (LLaVA-OV-7B, −77% tokens). At an aggressive 77% token reduction, CoVeR spreads its budget across every chair in the room, while VisPruner and SeGPruner cluster tokens on just a few patches and undercount. Asked "How many chairs are to the left of the table?", the full-token model correctly answers Four. VisPruner concentrates on near-duplicate views of one region and answers Zero; SeGPruner still misses most chairs and answers One. CoVeR spreads its budget across the whole scene and matches the full-token answer, Four — with 5.1× fewer TFLOPs, 1.9× faster inference, and 4.3× smaller KV cache.

BibTeX

@misc{bui2026covercoveragebasedtokenpruning,
  title         = {CoVeR: Coverage-Based Token Pruning for Multi-View 3D Reasoning in VLMs},
  author        = {Bui, Nhat-Tan and Elangovan, Varshini and Anugu, Arun Reddy and Mohan, Sreyas and Ye, Wei and
                   Wang, Dilin and Huang, JQ and Ranjan, Rakesh and Chharia, Aviral and De la Torre, Fernando},
  year          = {2026},
  eprint        = {2609.08345},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2609.08345}
}