MambaLIFT: Reference-Based Mechanism-Aware Attribution for VMamba
Organizations: School of Electrical and Computer Sciences, Indian Institute of Technology Bhubaneswar, Bhubaneswar 752050, India
Abstract
Attributing predictions of visual state-space models remains challenging because input-dependent selective dynamics couple feature selection, recurrent state propagation, and shared computational pathways. We introduce MambaLIFT, a reference-based, multiplier-first attribution framework for VMamba that propagates contributions through its computational graph using interaction-aware rules for selective-scan injection, recurrent memory, LayerNorm, and shared pathways. MambaLIFT decomposes the reference-relative class-logit difference, , by spatial position, VSS block, scan direction, and computational mechanism. Experiments on ImageNet-1K evaluate VMamba-T and VMamba-B, with hyperparameters selected on a calibration set and faithfulness assessed on a disjoint held-out evaluation set. MambaLIFT achieves numerical conservation to floating-point precision across audited attribution interfaces in both configurations. Under mean-colour pixel flipping, it achieves significantly higher mean faithfulness scores than MambaLRP on both configurations and the highest scores among the evaluated methods, including Integrated Gradients with matched references. Under blurred replacement, MambaLIFT is non-inferior to MambaLRP on VMamba-T and equivalent on VMamba-B, indicating that its advantage depends on the replacement protocol. Ablations on VMamba-T show that interaction allocation, temporal propagation, and pathway composition affect attribution maps. Mechanism-level analyses further show that exact relevance conservation does not imply additive effects under finite interventions: the mechanism decomposition accounts for the reference-relative logit difference but does not estimate intervention effects or establish causal validity.
Figures & tables
| Method | Model / setting | What is explained | Output / analytical resolution |
|---|---|---|---|
| MambaLRP [ 4 ] | Mamba, Vision Mamba | Class logit through layer-wise relevance propagation | Relevance per input token or pixel; selective-state-space propagation |
| LaTIM [ 5 ] | Mamba | Token-to-token interactions within Mamba representations | Token interaction maps |
| Hidden attention [ 6 ] | Mamba, Vision Mamba | Selective scan viewed as implicit attention | Attention-like influence maps |
| Controllability analysis [ 7 ] | Vision State-Space Models | Influence of input positions on internal state dynamics | Jacobian/Gramian-based influence scores at patch/state level |
| MambaLIFT | VMamba-T/B | Reference-relative class-logit difference | Relevance per spatial position, block, scan direction, and computational mechanism through the selective SS2D graph |
| Component | Setting | Parameters / Protocol |
|---|---|---|
| Models | VMamba-T, VMamba-B | vssm1_tiny_0230s , vssm_base_0229 ; , no output gate |
| Data | ImageNet-1K | Resize 256, centre crop ; 1,000 calibration + 1,000 held-out |
| Baselines | GI, SmoothGrad, IG, LRP-LN, MambaLRP | SmoothGrad: 30 samples, ; IG: 30 steps; LRP-LN: LN-rule + LRP-0; MambaLRP: VMamba adaptation, |
| MambaLIFT | Reference-based | uniform-noise references; ; Gaussian-blur reference ( pixels) used in reference-sensitivity analysis |
| Faithfulness | Pixel flipping | MoRF/LeRF, 10% steps; ranking; mean-colour replacement; Gaussian-blur replacement ( pixels) |
| Statistics | Bootstrap + Wilcoxon | 10,000 percentile-bootstrap resamples; paired two-sided Wilcoxon; Holm correction; margin |
| Evaluation | VMamba-T | VMamba-B |
|---|---|---|
| Audited interfaces | 138 | 194 |
| Worst relative audit error | ||
| MambaLIFT, target | rel. ; | rel. ; |
| GI, target | rel. ; | rel. ; |
| MambaLRP, target | rel. ; | rel. ; |
| Method | 95% CI | Median | (%) | |||
| VMamba-T | ||||||
| Random | 49.9 | |||||
| GI | 51.1 | |||||
| SmoothGrad | 80.1 | |||||
| IG, mean-colour baseline | 58.0 | |||||
| IG, noise baseline ( ) | 97.2 | |||||
| vs. MambaLRP | vs. IG, noise baseline ( ) | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Replacement | Model | Difference | 95% CI | 90% CI | Outcome | Difference | 95% CI | ||
| Mean colour | VMamba-T | Superior | |||||||
| Mean colour | VMamba-B | Superior | |||||||
| Blur, | VMamba-T | Non-inferior | |||||||
| Blur, | VMamba-B | Equivalent | |||||||
| VMamba-T | VMamba-B | |||
| Method | s/image | GB | s/image | GB |
| Forward pass | 0.012 | 0.02 | 0.018 | 0.05 |
| GI | 0.027 | 0.15 | 0.040 | 0.35 |
| SmoothGrad (30) | 0.205 | 1.21 | 0.522 | 3.22 |
| IG (30) | 0.191 | 1.18 | 0.510 | 3.20 |
| LRP-LN | 0.037 | 0.12 | 0.055 | 0.28 |
| Allocation | rel. | |||
|---|---|---|---|---|
| Direct | 1 | 1 | 1 | 0 |
| Nested | 1 | 0.9955 | 0.9256 | 0.05927 |
| Nested | 1 | 0.9968 | 0.9361 | 0.04946 |
| Nested | 1 | 0.9948 | 0.9226 | 0.06161 |
| Mean of nested a | 1 | 1.0000 | 1.0000 | |
| Seq. | 1 | 0.8254 | 0.6076 | 0.4336 |
| Ablation | Retained pathway | |||
|---|---|---|---|---|
| Complete | All pathways | 1.000 | 1.000 | 1.000 |
| Step size | Injection | 2.443 | 0.935 | 0.722 |
| Transition | 0.352 | 0.763 | 0.513 | |
| Content | Recurrent injection | 0.264 | 0.716 | 0.517 |
| -skip | 0.510 | 0.753 | 0.534 | |
| Scan fan-out | Content | 0.916 | 0.904 | 0.656 |
| Variant | rel. | |||
|---|---|---|---|---|
| Midpoint | 1 | 1 | 1 | 0 |
| Carry | 1 | 0.9371 | 0.7354 | 0.2313 |
| Carry | 1 | 0.5420 | 0.4153 | 1.259 |
| Window | 0.7132 | 0.7819 | 0.5424 | 0.5402 |
| Window | 0.6810 | 0.8101 | 0.5853 | 0.5143 |
| Window | 0.6156 | 0.8634 | 0.6551 | 0.4942 |
| Reference | |||
|---|---|---|---|
| Mean colour | 8.944 | 1.0000 | 1.0000 |
| Black | 8.928 | 0.1085 | 0.2819 |
| White | 8.983 | 0.2157 | 0.3240 |
| Blur, | 5.283 | 0.1111 | 0.1257 |
| Blur, | 7.756 | 0.1940 | 0.1634 |
| Blur, | 8.529 | 0.2977 | 0.2121 |
| Mechanism | Noise | Mean colour | Blur |
|---|---|---|---|
| Local: Spearman of and | |||
| Memory | |||
| Selection | |||
| alone | 0.25 | 0.19 | 0.02 |
| alone | 0.13 | 0.15 | 0.30 |
| alone | |||
| MambaLIFT, reference | ||||
|---|---|---|---|---|
| Mean colour | Blur, | Noise, | MambaLRP | |
| 0 | 2.0453 | 2.7639 | 2.7888 | 1.8912 |
| 0.15 | 3.7525 | 3.7447 | 4.0976 | 3.7907 |
| 0.25 | 3.8153 | 3.7059 | 4.0677 | 3.8475 |
| 0.30 | 3.8310 | 3.6642 | 4.0667 | 3.8362 |
| 0.35 | 3.8089 | 3.6316 | 4.0690 | 3.8181 |
Feature Attribution
Feature attribution estimates which input features or internal components contribute to a model’s predictions.
See all 258 papers