sae-ablation-clamping-zero-method
IN premise — summaries/2026/08/24/templeton-2024-scaling-monosemanticity-chunk-6.md
Created 2026-08-25T02:58:38+00:00
Feature ablation is performed by clamping a feature's activation to zero at a specific token position during a forward pass, measuring the full nonlinear causal effect on logit difference.
Summary
To test what a specific feature actually does, the system pins that feature's output to zero at one position in the input and watches how the model's final answer changes, capturing the full ripple effect through every later layer rather than just a local, linear estimate. This defines the ground-truth measurement protocol, so every downstream ablation result is interpreted against the same causal standard.