Capability-preserving neuron steering improves language model behavior control
CANS: Capability-Preserving Neuron Steering for Large Language Models
Artificial Intelligence
Summary
Controlling how large language models behave can sometimes reduce their overall usefulness. The authors present CANS, a way to steer a model’s responses toward desired behaviors without losing its general abilities. Their method adjusts parts of the model based on mathematical shortcuts, avoiding extra training or slower runs. Tests show CANS can reduce unsafe or incorrect outputs better while keeping the model's skills intact.
What this means in practice
- •For ai safety engineers: Control harmful outputs in large language models with minimal impact on overall model performance.
- •For content moderation teams: Reduce toxicity and hallucination in AI-generated text while preserving language understanding skills for general tasks.
- •For automated customer support developers: Improve model responses for specific behavior requirements without retraining or slowing down response time.$Commercial implications: Enables safer and more reliable chatbots better meeting company standards, which can be marketed as higher-quality AI support tools.
Authors
Yucheng Wu, Chengyi Xie, Chenpeng Wang, Liangming Pan
Abstract
Activation steering offers training-free and dynamic control of target behaviors in large language models, but existing methods can impair general capabilities and require domain-specific intervention designs. We introduce Capability-Preserving Neuron Steering (CANS), a unified framework that balances target-behavior control with capability preservation. To make target-behavior improvement and general-capability preservation tractable, we introduce two proxies: log-probabilities of target tokens in model generated responses, and output-distribution KL divergence on general-purpose inputs. Combining the local approximation of the two proxies yields a closed-form solution, enabling lightweight steering without iterative model training or additional inference-time computation. The resulting neuron-wise scaling coefficients achieve effective steering while limiting general-capability loss. A shared data-construction and intervention procedure supports adaptation across domains using sampled responses with target-behavior filtering and target-span annotations. Across safety, toxicity, and hallucination control on Qwen and Llama models, CANS achieves comparable or better steering while better preserving MMLU, BBH, and IFEval performance than SFT, CAA, and neuron-level baselines. On Qwen3-4B, CANS reduces HarmBench attack success from $15.50\%$ to $1.75\%$ with a $1.11$-point decrease in average capability. These results support CANS as a practical approach to behavioral control across domains.