Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.
MANCE is our new method for concept erasure. @matan_avitan_'s idea is very simple: estimate the manifold using natural input representations, and make sure that any update you make stays on that manifold. It works very well in practice, 1/ https://twitter.com/matan_avitan_/status/2076994007905751256
both as an add-on to other erasure methods (INLP, LEASE, etc.), and on its own. We perform extensive experiments (119 settings), both on text and vision, multiple concepts, and LLMs/VLMs. We show how concepts such as sycophancy, gender, safety, and a variety of 2/
vision concepts can be surgically removed from models' activations, while minimizing the harm to other concepts. You should check out the paper! https://arxiv.org/abs/2607.03973
MANCE is our new method for concept erasure. @matan_avitan_'s idea is very simple: estimate the manifold using natural input representations, and make sure that any update you make stays on that manifold. It works very well in practice, 1/ https://twitter.com/matan_avitan_/status/2076994007905751256
both as an add-on to other erasure methods (INLP, LEASE, etc.), and on its own. We perform extensive experiments (119 settings), both on text and vision, multiple concepts, and LLMs/VLMs. We show how concepts such as sycophancy, gender, safety, and a variety of 2/
vision concepts can be surgically removed from models' activations, while minimizing the harm to other concepts. You should check out the paper! https://arxiv.org/abs/2607.03973
Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.