cs.CVNov 14, 2026

Size-Matched Nulls Reveal What Block Ablation Measures in MambaVision

Authors: Tahmina Kousar, Tehseen Zia

Organizations: Department of Computer Science, COMSATS University Islamabad, Islamabad, Pakistan

Abstract

Hybrid vision backbones interleave Mamba state-space mixers with self-attention, and a block's role is usually probed by removing it and measuring the damage. We test what that measurement means in MambaVision. On all 50,000 ImageNet validation images and three model sizes, every block ablation is paired with a size-matched null, the same perturbation with its channels shuffled or their signs flipped. In top-1 accuracy and in true-class rank, real and null damage are nearly indistinguishable in absolute terms: pooled over the blocks of MambaVision-T, the content effect lies within [-0.044,+0.016] percentage points (90% interval). Two attention-only controls show that the test detects content, with real ablations worse than their nulls by up to 4.2 points in DeiT-Small and 1.0 point in Swin-T. Part of MambaVision's near-equivalence reflects how little damage its blocks do (0.11--0.34 points on average, against 1.7 and 2.1). Scaled by damage, its content share is comparable to Swin-T's in accuracy, lower in rank, and far below DeiT-Small's: single-block damage in MambaVision is small and mostly positional. Ablation also hides what blocks carry together: removing all blocks of one type compounds damage to 3.7--8.4 times the sum of single effects, far more than size-matched nulls, including nulls that keep the alignment between the blocks' perturbations; for adjacent pairs, aligned nulls reproduce most of it. Freezing input-dependent selection in all Mamba mixers shows that it accounts for 20--32% of the Mamba contribution, mainly through the B and C projections. We turn these results into a protocol for causal attribution in hybrid networks.

Figures & tables

Explore similar work

Jul 21, 2026

Norm or Direction? Decoding Vision Mambas for High-Resolution Vision

Vision Mamba models replace quadratic self-attention with linear complexity selective state space models (SSMs), emerging as efficient visual backbones. However, MambaOut demonstrates that a Gated CNN block can match or exceed VMamba on image classification, questioning the necessity of SSMs for vision. This raises a fundamental question: do VMamba and MambaOut encode visual information differently at the representation level? To investigate, we apply cross model centered kernel alignment (CKA) analysis and find that VMamba's final stage blocks form representations distinctly different from both MambaOut and its own preceding blocks. We therefore focus on the final block features, decomposing each spatial token into magnitude and direction. MambaOut concentrates class-discriminative information in high-norm foreground tokens that align with Grad-CAM attribution. VMamba, by contrast, produces high-norm tokens predominantly in background regions, misaligned with Grad-CAM, yet preserves discriminative signals primarily in token directions. These observations reveal that the two models rely on different encoding strategies. We connect this difference to high-resolution classification and semantic segmentation. VMamba distributes logit support broadly across object regions, whereas MambaOut relies on sparse dominant tokens, a strategy that becomes less stable as token counts grow. Under full fine-tuning for segmentation, VMamba consistently outperforms MambaOut. These results suggest that VMamba's advantage in dense prediction stems not merely from the SSM mechanism or sequence length, but from how semantic evidence is organized across token magnitude, direction. Ultimately, we conclude that token magnitude and directional structure serve as critical axes for improving visual backbones, particularly under dense supervision.
Oct 8, 2026

Bridging Human and Model Attention: Explainability Analysis of CNN, Mamba, and ViT Architectures with Gaze-Based Validation

High-risk applications of deep learning require precise characterization of model decision-making. We conduct a systematic explainability evaluation across Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Vision Mamba (ViM). Unlike predominantly qualitative studies, we quantify the alignment between model attention and human visual perception using eye-tracking ground truth. We formalize this comparison via a standardized normalization pipeline mapping post-hoc attribution maps and gaze heatmaps onto spatial probability distributions. Alignment is evaluated across fine-grained classification, medical imaging, and emotion recognition using architecture-specific XAI methods. We measure fidelity using Jensen-Shannon similarity, Chi-square similarity, and Pearson correlation, three complementary estimators capturing support overlap, mass discrepancy, and structural co-variation. Our results reveal distinct attention dynamics: CNNs tend to exhibit a more global focus, with spatially more diffuse saliency maps that correlate well with human perception in general contexts, while ViTs demonstrate fine-grained, patch-level attention, often focused on discriminative microfeatures. The Mamba architecture shows biased behavior: it achieves strong explainability in medical tasks but exhibits divergent intensity distributions in natural images; our analytical characterization of the metrics makes this dissociation explicit, since near-unit support overlap is attained simultaneously with a vanishing linear correlation. We further analyze the effect of backbone freezing and ensemble strategies on model interpretability. The results highlight how architectural and training choices influence the alignment between model saliency and human visual attention.
Aug 4, 2026

A Theory of Conditional Collapse under Low-Rank Weight-Space Ablations: I. The Single-Block Theory and Synthetic Validation

Activation patching and weight-space ablation both claim a component is causally responsible for a behavior, yet they act on different objects: one forward pass versus the parameters behind every forward pass. We ask when they agree. We study an idealized model where a conditional computation is carried additively through a residual stream, F(x)=F0(x)+∑iαi(x)viF(x)=F_0(x)+\sum_iα_i(x)v_i, read out by a linear functional, and prove three exact results. First, deleting a subset of carriers collapses a matched input pair onto the same unconditional output if and only if the removal is symmetric on the pair and leaves no outside contrast; the error is deterministic, and we give its exact form even when the two conditions hold only approximately. Second, patching a carrier moves the readout by its donor-receiver contrast, while ablating it moves the readout by its absolute level; neither bounds the other, and we construct pairs where every single-carrier patch flips the decision while no single-carrier ablation does. Third, for an attention head composed with its own layer's normalization and MLP, we derive an exact first-order interaction formula with a provably second-order remainder, vanishing identically when only the MLP is ablated but not, in general, when a head is. Small transformers trained on a synthetic conditional task illustrate all three predictions: across thirty-nine ablation configurations the measured interaction is strongly rank-correlated with the idealized model's predictive accuracy (Spearman −0.83-0.83), and a second task and architecture reproduces the same pattern, including a further polarity reversal. The single-block interaction result extends past one residual block, and the synthetic validation is tested against a real pretrained model, in a companion paper that takes this theory further along both axes.