Nous Research drops Contrastive Neuron Attribution method

CNA is a major step forward for model steering, moving past noisy "activation additions" toward surgical neuron manipulation. It provides a way to "lobotomize" safety guardrails or enhance specific capabilities by touching only 0.1% of the model. The method identifies behavioral circuits like the "refusal gate" using only forward passes, bypassing the need for expensive Sparse Autoencoders. It reveals that alignment fine-tuning crystallizes existing discrimination features found in base models rather than creating new structures. The toolkit is available as the neural-steering repository on GitHub, offering stability at high intervention strengths where previous methods caused model collapse.