Copy page View as Markdown

Tutorials

• Logit lens: where in depth a prediction emerges.
• Attention patterns: scoring attention heads.
• Contribution decomposition: direct logit attribution over the standard contributions.
• Ablation: the effect of removing a contribution, a head or a block.
• Activation patching: a value from a clean run patched into a corrupt one.
• Steering: a direction added to the residual stream.
• Probing: a linear probe per layer.
• Cross-family sweep: one experiment over several checkpoints.
• Expert ablation: every expert's effect on a prediction.
• Image pathway: the image's effect in a vision-language model.
• DeltaNet state: the recurrent state of a hybrid, read and patched.