Mechanistic Interpretability // Layer-by-Layer Probing

What Does Each Layer Look At & Infer?

Step through the transformer's 12 blocks from shallow to deep. Observe how attention transitions from raw textures to semantic parts, and see what the model infers at every depth.

⚡ Transformer Depth Controller (12 Sequential Blocks)
Selected: Layer 6
L0 (Input) L11 (Output)
Mid Bottleneck
1. Layer Attention Overlay: 60%
2. Saccade Tokens Top 12 Patches
Visual Token Budget (K): 12 / 196 (6.1% of image)
Heatmap Blend Intensity:
Inference at this Depth Probed @ L6
Target Class Progression (L0 ➔ L11)
Intrinsically Interpretable Guarantee

Unlike post-hoc heatmaps (Grad-CAM), the downstream classification was computed exclusively using the selected patches.

Information Discarded: 93.9% of pixels
Gradient Isolation: 0.000000 (No Leakage)
Attended Patches Breakdown (Ranked by Importance in Layer 6) Hover patch to locate