CANS: A Capability-Preserving Neuron-Oriented Method for Large Language Models

The researchers propose CANS, a neuron-oriented framework that balances target behavior control with general capability retention, enabling lightweight behavioral intervention without iterative training or additional inference computation. The method uses the log probability of target tokens in model-generated responses and the KL divergence of output distributions on general inputs as proxy metrics, combines local approximation to obtain a closed-form solution, and generates neuron-level scaling coefficients; the shared data construction and intervention pipeline adapts to different domains through target behavior filtering and target segment annotation. On safety, toxicity, and hallucination control tasks for Qwen and Llama models, CANS's steering effectiveness is comparable to or better than SFT, CAA, and neuron-level baselines, while better preserving MMLU, BBH, and IFEval performance. On Qwen3-4B, CANS reduces the HarmBench attack success rate from 15.50% to 1.75%, with an average capability drop of 1.11 points.