neuron-ablation

NousResearch releases Contrastive Neuron Attribution (CNA), a method to steer LLM behavior by ablating sparse MLP circuits without training autoencoders or degrading benchmarks, validated on refusal circuits across models up to 70B parameters.