Sensory stimuli evoke spiking neural responses that innately or after learning drive suitable behavioral outputs. How are these spiking activities intrinsically patterned to encode for innate preferences, and could the neural response organization impose constraints on learning? We examined this issue in the locust olfactory system. Using a diverse odor panel, we found that ensemble activities both during (‘ON response’) and after stimulus presentations (‘OFF response’) could be linearly mapped onto overall appetitive preference indices. Although diverse, ON and OFF response patterns generated by innately appetitive odorants (higher palp-opening responses) were still limited to a low-dimensional subspace (a ‘neural manifold’). Similarly, innately non-appetitive odorants evoked responses that were separable yet confined to another neural manifold. Notably, only odorants that evoked neural response excursions in the appetitive manifold could be associated with gustatory reward. In sum, these results provide insights into how encoding for innate preferences can also impact associative learning.