Sheaf method isolates paraphrase-stable content in transformers
A 20-dimensional content subspace beats LEACE by a mean 17.1 points on held-out CounterFact retrieval across eight architectures.
Stories per day · 30 days
Relatedcalibration1concept-erasure1counterfact1domain-shift1interpretability1minimax1paraphrase1representation-repair1sheaf-theory1sleep-eeg1transformers1
2 stories, newest first · one card per story
Stories tagged #leace
Detecting a Shift Is Not Enough: Exact Minimax Limits of Linear Representation Repair
Exact minimax limits show linear shift repair needs far more calibration signal than detection.