@NousResearch: To check that CNA isolates only the intended behavior, we evaluate steered models on MMLU across a range of steering st…

X AI KOLs Following Papers
steering neuron-attribution ablation llm interpretability mlp

Summary

Nous Research released Contrastive Neuron Attribution (CNA), a method to steer LLM behavior by identifying and ablating sparse circuits in MLP neurons without training sparse autoencoders or degrading general benchmarks, validated on multiple large language models.

Today we release Contrastive Neuron Attribution (CNA), a method for steering LLM behavior by identifying and ablating sparse circuits in the MLP basis without training a sparse autoencoder, modifying weights, or degrading general capability benchmarks. Given a small set of contrastive prompt pairs that elicit a target behavior and its opposite, CNA isolates the top 0.1% of MLP neurons whose activations differ most between the two sets. Ablating that small circuit removes the behavior while leaving the rest of the model intact, and the intervention remains robust at high strengths where residual-stream methods like Contrastive Activation Addition (CAA) start to degrade. Validated on the refusal circuit across 8 instruct-tuned models, including Llama-3.1-70B, Llama-3.2-3B, Qwen2.5-72B, and Qwen2.5-14B. The work on CNA was led by @yaboilyrical, with support from @qorprate and @karan4d.
Original Article
View Cached Full Text

Cached at: 05/19/26, 04:50 PM

Today we release Contrastive Neuron Attribution (CNA), a method for steering LLM behavior by identifying and ablating sparse circuits in the MLP basis without training a sparse autoencoder, modifying weights, or degrading general capability benchmarks.

Given a small set of contrastive prompt pairs that elicit a target behavior and its opposite, CNA isolates the top 0.1% of MLP neurons whose activations differ most between the two sets. Ablating that small circuit removes the behavior while leaving the rest of the model intact, and the intervention remains robust at high strengths where residual-stream methods like Contrastive Activation Addition (CAA) start to degrade.

Validated on the refusal circuit across 8 instruct-tuned models, including Llama-3.1-70B, Llama-3.2-3B, Qwen2.5-72B, and Qwen2.5-14B.

The work on CNA was led by @yaboilyrical, with support from @qorprate and @karan4d.

Similar Articles

Targeted Neuron Modulation via Contrastive Pair Search

Hugging Face Daily Papers

Contrastive neuron attribution (CNA) identifies a sparse set of MLP neurons that distinguish harmful from benign prompts, enabling effective behavioral steering in instruction-tuned LLMs without degrading output quality. The method reduces refusal rates by over 50% on jailbreak benchmarks while preserving fluency.

Multi-Attribute Steering of Language Models via Targeted Intervention

arXiv cs.CL

MAT-Steer introduces a novel inference-time intervention framework for steering LLMs across multiple conflicting attributes by learning sparse, orthogonal steering vectors that selectively target tokens relevant to each attribute, achieving gains in QA tasks and generative tasks over prior methods.

Natively Unlearnable Large Language Models

arXiv cs.LG

The paper proposes NULLs (Natively Unlearnable LLMs), a model class that isolates source-specific contributions in sparsely activated sinks while sharing backbone neurons, enabling clean unlearning of individual data sources without retraining and preserving general language capabilities.