NeuronEye is a framework for vision-language models that constructs a sparse, concept-level neuron vocabulary to selectively activate query-relevant visual concepts during inference.