Spatial-Aware Reduction Framework: Towards Efficient and Faithful Visual State Space Models

Researchers introduce STORM, a spatial-aware token reduction framework that addresses performance collapse in visual state space models like Mamba when applying token reduction techniques. By maintaining structural integrity and two-dimensional grid topology during compression, STORM achieves significant accuracy recovery, particularly on VMamba with up to 63.3% improvement while operating as a training-free plug-and-play module.

The advancement of efficient visual processing models has encountered a critical technical bottleneck. Mamba-based architectures demonstrate strong efficiency in handling long visual sequences, yet existing token reduction methods cause severe performance degradation when applied to structurally enhanced variants. The root cause lies in a fundamental architectural mismatch: conventional reduction techniques ignore spatial relationships, breaking the two-dimensional structural assumptions that selective scanning mechanisms depend upon.

STORM addresses this gap by reformulating token reduction as a structured operation on spatial units rather than treating tokens as an unordered collection. The framework enforces localized constraints that preserve both grid topology and neighborhood coherence, effectively treating visual data as inherently spatial rather than sequential. This represents a paradigm shift in how reduction methods interact with vision models.

The practical implications are substantial for developers and researchers optimizing vision transformers. The training-free nature of STORM as a plug-and-play module means immediate applicability across existing pipelines without requiring model retraining or extensive computational investment. Results demonstrate state-of-the-art pruning accuracy across diverse Mamba backbones, with VMamba recovery reaching 63.3% improvement and PlainMamba maintaining near-ViT parity with only 1.0% accuracy loss.

This work signals a broader trend in AI optimization: generic reduction strategies fail when models encode structural assumptions. Future model compression research will likely prioritize architecture-aware methods that respect underlying geometric and topological properties. For AI practitioners, STORM provides an immediate tool for deploying efficient vision models without sacrificing accuracy, potentially accelerating adoption of state space models in resource-constrained environments.

• →STORM enables training-free token reduction while maintaining model performance through spatial-aware constraints on grid topology
• →VMamba achieves 63.3% accuracy improvement over prior reduction methods using the STORM framework
• →Existing token reduction techniques fail because they ignore two-dimensional structural requirements of selective scanning mechanisms
• →The plug-and-play module design allows immediate integration into existing reduction pipelines without retraining
• →PlainMamba maintains comparable performance to ViT with only 1.0% accuracy degradation under STORM compression