CoVeR Retains 93.5% of 3D VLM Accuracy Using 8% of Tokens

CoVeR uses only about 8% of visual tokens and keeps 93.5% of a multi-view 3D VLM's full-token performance, according to a paper posted on Hugging Face. The pruner is training-free and deterministic, and picks tokens from their coordinates alone with no learned signals.

The abstract frames the setup as a coverage problem. Rendering a 3D scene as many views produces "thousands of redundant visual tokens whose cost grows with every view." Existing pruners fall into two families. Learned-importance methods, the authors write, "keep near-duplicate tokens from a few prominent regions and leave most of the scene unrepresented." Voxelization methods, meanwhile, "cannot enforce an exact token budget and saturate as multi-view observations overlap in 3D."

CoVeR runs in two stages. An adaptive voxel grid keeps one representative token per occupied voxel; farthest-point sampling then expands the selection until an exact per-scene budget is filled. The paper evaluates on ScanQA, SQA3D and OpenEQA using LLaVA-OneVision-7B, Video-3D LLM, Qwen2.5-VL-7B and Qwen3-VL-8B, with a single coverage weight a=0.4 kept fixed across all four models. On the same setup it reports beating prior state of the art "by 3.9 percentage points" on average.

At a 9% token budget the authors report 13.3x fewer LLM TFLOPs, 10.7x smaller KV cache, 2.9x inference speedup, and a 1.1% relative accuracy drop. It slots into a busy run of inference efficiency work we have been tracking, our 93rd inference piece in the last 90 days. The abstract publishes no per-model latency comparisons against the base runs, and the Hugging Face page shows no author affiliations.