Nous Research Releases Contrastive Neuron Attribution (CNA): Sparse MLP Circuit Steering Without SAE Training or Weight Modification

Nous Research introduced contrastive neuron attribution (CNA), a neuron-level method for finding and steering sparse refusal circuits in instruction-tuned LLMs.

By comparing harmful and benign prompts, CNA identifies specific MLP neurons; ablating only about 0.1% of activations cut refusal rates by more than half in most tested models.

The method worked across Llama 3.1, Llama 3.2, and Qwen 2.5 while preserving output quality.

The study also suggests the relevant late-layer structure exists in base models before fine-tuning, and alignment training repurposes it into a refusal gate.

CNA offers a precise alternative to coarse steering vectors and heavier SAE-based approaches, without extra training or weight changes.