loading scene data…
Forget, or just suppress?
scene
induction
refusal
unlearning
emergence
SAE features
attribution graph
unlearned checkpoint
intervention
baseline
ablate refusal
ablate forget
metric
answer probability
generation ROUGE
knows the answer
forgot it
gauge = mean score for this checkpoint
full ceiling ·
retain90 floor
baseline ghost ring shows the jump under ablation
drag = orbit · wheel = zoom
❚❚
0.5×
1×
2×
4×