refusal-aliases

The paper introduces AMRA, a weight-editing method to mitigate abliteration in large language models by obscuring the refusal signal, improving post-abliteration refusal scores on Llama-3-8B and Gemma-2-9B with minimal utility degradation.