X Vmamba Controllability Eval

---
name: x-vmamba-controllability-eval
description: Probes the spatial feature flow and patch-level influence in Vision Mamba models using classical control theory. It quantifies how input image patches drive hidden state dynamics across hierarchical layers, revealing domain-specific diagnostic feature extraction patterns. Use when the user wants to benchmark on CMMD, DermaMNIST, BloodMNIST, or asks about evaluating this task. Reports influence score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2511.12694
  bibtex_key: mabrok2025xvmamba
  confidence: medium
---

# x-vmamba-controllability-eval

> X-VMamba: Explainable Vision Mamba — Mabrok et al. (2025) (arXiv:2511.12694, 2025)

## What this evaluates

Probes the spatial feature flow and patch-level influence in Vision Mamba models using classical control theory. It quantifies how input image patches drive hidden state dynamics across hierarchical layers, revealing domain-specific diagnostic feature extraction patterns.

## Datasets

- **CMMD** — total ?; splits: train (-1), test (-1)
- **DermaMNIST** — total ?; splits: train (-1), test (-1)
- **BloodMNIST** — total ?; splits: train (-1), test (-1)

## Metrics

- `influence score` **(primary)** — range: other
  - Jacobian-based metric quantifying the contribution of each input patch to the model's hidden state dynamics and final classification output. Computed via the Jacobian of the hidden states with respect to input patches, aggregated to produce per-patch influence scores that evolve across network layers.

## Input / output format

**Input**: Medical images (grayscale mammography, RGB dermatology, or microscopic blood cell images) fed into the MedMamba architecture.

**Output**: Per-patch controllability heatmaps and influence score distributions across the first three hierarchical layers, indicating spatial selectivity and diagnostic feature localization.

## Scoring recipe

```python
def compute_influence_scores(model, image):
    # Forward pass to obtain hidden states H
    H = model(image)
    # Compute Jacobian of hidden states w.r.t input patches
    J = jacobian(H, image)
    # Aggregate Jacobian norms to derive per-patch influence scores
    raw_scores = torch.norm(J, dim=[...]).mean(dim=0)
    # Normalize scores to enable cross-layer comparison
    normalized_scores = raw_scores / raw_scores.max()
    return normalized_scores
```

## Common pitfalls

- Controllability heatmaps are derived from Jacobian/Gramian control theory, not standard attention mechanisms; they reflect state-space dynamics rather than softmax weights.
- Influence scores are highly sensitive to the image scanning strategy (e.g., directional sequencing in VMamba); results cannot be generalized across different scanning schemes without explicit comparison.

## Evidence (verbatim from paper)

> We evaluated to Jacobian-based approach as it is more broad and doesn’t contain the assumption of Gramian-based model, the diagonal $A$ matrix. We evaluated the controllability pattern evolution, influence score distribution, and more detailed analysis of influence score for different image scanning scheme in Mamba models.

## Citation

```bibtex
@misc{mabrok2025xvmamba,
  title={X-VMamba: Explainable Vision Mamba},
  author={Mabrok et al. (2025)},
  year={2025},
  note={arXiv:2511.12694}
}
```

- arXiv: 2511.12694