Probing Length Generalization in Mamba via Image Reconstruction
This addresses a specific limitation in Mamba for sequence modeling, offering insights for architectural improvements, but it is incremental as it builds on existing analysis.
The paper investigates Mamba's performance degradation when inference sequence lengths exceed training lengths, using an image reconstruction task to show that Mamba adapts to training lengths and fails to generalize beyond them, with a proposed variant improving performance across training lengths.
Mamba has attracted widespread interest as a general-purpose sequence model due to its low computational complexity and competitive performance relative to transformers. However, its performance can degrade when inference sequence lengths exceed those seen during training. We study this phenomenon using a controlled vision task in which Mamba reconstructs images from sequences of image patches. By analyzing reconstructions at different stages of sequence processing, we reveal that Mamba qualitatively adapts its behavior to the distribution of sequence lengths encountered during training, resulting in strategies that fail to generalize beyond this range. To support our analysis, we introduce a length-adaptive variant of Mamba that improves performance across training sequence lengths. Our results provide an intuitive perspective on length generalization in Mamba and suggest directions for improving the architecture.