SwordBench benchmarks steering methods for concept removal in vision models and shows that linear SVMs achieve strong separability and orthogonality but incur collateral damage, while sparse autoencoders often perform better and no method reaches perfect steering even in simple cases.
Nature , volume=
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
years
2026 2verdicts
UNVERDICTED 2representative citing papers
PUICL is a transformer pretrained on synthetic PU data from structural causal models that solves positive-unlabeled classification via in-context learning without gradient updates or fitting.
citing papers explorer
-
SwordBench: Evaluating Orthogonality of Steering Image Representations
SwordBench benchmarks steering methods for concept removal in vision models and shows that linear SVMs achieve strong separability and orthogonality but incur collateral damage, while sparse autoencoders often perform better and no method reaches perfect steering even in simple cases.
-
In-Context Positive-Unlabeled Learning
PUICL is a transformer pretrained on synthetic PU data from structural causal models that solves positive-unlabeled classification via in-context learning without gradient updates or fitting.