---
title: "CoverPrune: Coverage-Driven Token Pruning for 3D VLMs via Optimal Transport"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.13226v1"
permalink: "https://feed.astralor.com/items/itm_d6d7376d7de437b40475b9e9"
published: "2026-08-13T13:29:15.000Z"
score: 72
channel: "tech"
tags: ["3D VLM","arXiv","CoverPrune","Token Pruning","Optimal Transport","Spatial Reasoning"]
---

# CoverPrune: Coverage-Driven Token Pruning for 3D VLMs via Optimal Transport

## AI commentary

> The most compelling takeaway is the shift from “select diverse tokens” to “preserve evidence coverage,” which matters even more for 3D tasks than for 2D. For researchers working on 3D VLM inference acceleration or multimodal model compression, this coverage-centric angle could open a new path.

## At a glance

3D Vision-Language Models (3D VLMs) excel in spatial reasoning but suffer from massive visual token counts that create severe computational bottlenecks during inference. Existing token pruning methods mostly rely on diversity-based selection, which often drops representative prototype tokens in 3D scenes, breaking multi-view consistency and geometric structures. To address this, the team proposes CoverPrune, a training-free framework that formulates inference-time token pruning as an Optimal Transport problem, featuring Feature-Spatial-Temporal (FST) transport cost and a Spatial-Guided Greedy Selection (SGS) algorithm. An accelerated variant, CoverPrune-Lite, is also introduced. Extensive experiments on multiple 3D spatial reasoning benchmarks show state-of-the-art token efficiency while maintaining robust performance even under aggressive pruning budgets.

## Key points

- Proposes a paradigm shift for 3D VLM token pruning: from maximizing diversity to preserving visual evidence coverage.
- Introduces CoverPrune, a training-free framework built on Optimal Transport, plus an accelerated variant CoverPrune-Lite.
- Achieves state-of-the-art token efficiency across multiple 3D spatial reasoning benchmarks while maintaining robust performance under aggressive pruning budgets.

## Full text

While 3D Vision-Language Models (3D VLMs) have demonstrated remarkable spatial reasoning capabilities, they suffer from massive visual token counts that create severe computational bottlenecks during inference. Existing token pruning methods primarily rely on diversity-based selection, discarding similar tokens to maximize dispersion. However, in 3D environments, this approach frequently drops representative prototype tokens in favor of outliers, breaking the multi-view consistencies and geometric structures essential for spatial reasoning. In this paper, we propose a paradigm shift for 3D VLM token pruning: from maximizing diversity to preserving visual evidence coverage. We introduce CoverPrune, a training-free framework that formulates inference-time token pruning as an Optimal Transport (OT) problem. To overcome the intractable combinatorial subset selection inherent in this formulation, we design the Feature-Spatial-Temporal (FST) transport cost and target capacity, along with an efficient Spatial-Guided Greedy Selection (SGS) algorithm to approximate the OT objective. Furthermore, we propose CoverPrune-Lite, an accelerated variant utilizing spatially structured local matching for minimal overhead. Extensive experiments across multiple 3D visual-spatial reasoning benchmarks demonstrate that our methods achieve state-of-the-art token efficiency, maintaining robust reasoning performance even under highly aggressive pruning budgets. Visit our project website at https://github.com/Brucess/CoverPrune.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_d6d7376d7de437b40475b9e9