Nous Research Introduces CNA: Controlling LLM Behavior Without Retraining

Nous Research introduced the Contrastive Neuron Attribution (CNA) method — a breakthrough approach to controlling the behavior of large language models. The method enables identifying and disabling individual neural circuits in MLP layers without requiring model retraining and without modifying its weights.

What is CNA and How Does It Work

Contrastive Neuron Attribution is a technique for identifying and ablating (disabling) sparse neural circuits in a multi-layer perceptron (MLP) network. Each MLP layer in a model contains thousands of neurons, but only a small fraction of them are responsible for specific behaviors, characteristics, or capabilities of the model.

The CNA method uses contrastive analysis — it compares network activations on examples where the target behavior is clearly expressed and on examples where it is absent. This approach makes it possible to identify the exact neurons that are most sensitive to the emergence or disappearance of the behavior of interest.

Once identified, these neurons can be deactivated, and the model stops demonstrating the undesirable characteristic. The simplicity of the method lies in its elegance: there is no need for additional training; it is sufficient to conduct the analysis and block the signal from the identified neurons during inference.

Main Advantage: No Retraining and No Weight Modification

The traditional way to control LLM behavior requires either full retraining (fine-tuning with a large dataset) or the application of a sparse autoencoder (SAE) — an additional neural network that learns to identify sparse components in the model's activations. Both approaches require significant computational resources and time, and often result in slight performance degradation.

CNA is fundamentally different. The method requires no retraining and does not modify model weights at all. Behavior control occurs exclusively at the level of neural activations — they can simply be deactivated during inference. This makes the process much faster, cheaper, and importantly, completely reversible: if the solution doesn't work, you can simply return the neurons to their original state.

A key result of the Nous Research study confirms that applying CNA does not cause degradation of the model's overall performance. After using the method, the model retains:

• High results on standard benchmarks (MMLU, GSM8K, HumanEval)
• The full range of abilities unrelated to the target behavior
• Original inference speed and energy efficiency

Where This Can Be Applied

CNA is useful for removing or modifying undesirable model characteristics: biases in responses, toxic content, unwanted generation style, distorted associations. The method can also be applied to enhance desired capabilities — for example, to improve skills in specialized subject domains.

For organizations, this means the ability to adapt large public models (GPT-4, Claude, Llama) to their own requirements and values without the need for full retraining. This saves resources, accelerates deployment, and enables quick response to new requirements.

What This Means

CNA opens up a new way to fine-tune LLM behavior after their production release — cheaper and simpler than retraining, but far more effective than naive approaches like prompt engineering. This could significantly accelerate the development of safe, requirement-adapted AI systems, especially in regulated industries where model behavior is critical.