Nüwa: Mending the Spatial Integrity Torn by VLM Token Pruning
This addresses the issue of spatial integrity loss in VLMs for efficient processing, offering a novel solution that improves performance on visual grounding tasks, though it is incremental in the context of token pruning methods.
The paper tackles the problem of vision token pruning in Vision Language Models (VLMs), which causes substantial degradation in visual grounding tasks due to loss of spatial integrity, and proposes Nüwa, a two-stage pruning framework that achieves state-of-the-art performance on VQA benchmarks (e.g., from 94% to 95%) and substantial improvements on visual grounding tasks (e.g., from 7% to 47%).
Vision token pruning has proven to be an effective acceleration technique for the efficient Vision Language Model (VLM). However, existing pruning methods demonstrate excellent performance preservation in visual question answering (VQA) and suffer substantial degradation on visual grounding (VG) tasks. Our analysis of the VLM's processing pipeline reveals that strategies utilizing global semantic similarity and attention scores lose the global spatial reference frame, which is derived from the interactions of tokens' positional information. Motivated by these findings, we propose $\text{Nüwa}$, a two-stage token pruning framework that enables efficient feature aggregation while maintaining spatial integrity. In the first stage, after the vision encoder, we apply three operations, namely separation, alignment, and aggregation, which are inspired by swarm intelligence algorithms to retain information-rich global spatial anchors. In the second stage, within the LLM, we perform text-guided pruning to retain task-relevant visual tokens. Extensive experiments demonstrate that $\text{Nüwa}$ achieves SOTA performance on multiple VQA benchmarks (from 94% to 95%) and yields substantial improvements on visual grounding tasks (from 7% to 47%).