Topology-Aware Layer Pruning for Large Vision-Language Models
About
Large Language Models (LLMs) have demonstrated strong capabilities in natural language understanding and reasoning, while recent extensions that incorporate visual inputs enable them to process multimodal information. Despite these advances, Large Vision-Language Models (LVLMs) incur substantial computational and memory costs, hindering deployment in resource-constrained scenarios. Existing layer pruning methods typically rely on local similarity metrics or static proxy signals, failing to capture the global and dynamic evolution of representations across model depth, which often leads to the removal of transition-critical layers. To address this limitation, we propose a topology-aware layer pruning framework for LVLMs. Specifically, we represent layer wise hidden states as point clouds and models their evolution using \textit{simplicial complexes}. By leveraging \textit{zigzag persistent homology}, we quantify inter-layer topological consistency and enable adaptive pruning that preserves critical representational transitions. Extensive experiments on diverse multimodal benchmarks demonstrate that the proposed framework consistently outperforms existing pruning methods across a wide range of sparsity ratios. Our code is available at https://github.com/zpc456/TopoVLM.
Key insights
Withdrawn preprint: the authors' June 2026 notice acknowledges uncredited methodology, code and data. These notes describe the historical v1 manuscript; its novelty claims should not be relied on. Read the withdrawal notice
What problem did this paper address?
The paper addressed high computational costs in Large Vision-Language Models. It proposed a topology-aware layer pruning framework that modeled hidden states as point clouds and used zigzag persistent homology to identify structurally redundant layers.
How did it differ from prior pruning methods?
The withdrawn manuscript described tracking topology across layers instead of relying on local similarity. It reported 91.6% of baseline performance for LLaVA-NeXT at 50% sparsity, 0.7 points above TAMP; withdrawal prevents treating this as an established novelty claim.
What were the practical findings useful for?
The historical experiments reported LLaVA-NeXT latency falling from 105.4 to 60.3 ms and peak VRAM falling by 43% at 50% sparsity. These were manuscript benchmarks, with no documented production deployment and the withdrawal notice still applicable.
What limitations did the authors state?
The authors stated the framework did not consider hardware-aware codesign for energy efficiency, noting that reduced model depth may not proportionally lower energy consumption across diverse hardware platforms.