Open-ended agentic search for neural encoding models

Models trained for computer vision predict visual cortical responses well, but their complexity obscures the computations underlying this success. Here we introduce neuro-autoresearch, a framework in which language-model agents use neural feedback to construct compact, interpretable models. Agents iteratively revise explicit computational operations, guided by held-out neural predictivity, without training the models' internal weights. Independent searches on human fMRI and macaque neuronal recordings discover models that combine oriented filtering, color opponency, normalization and pooled image statistics. These models outperform all tested neural networks in early visual cortex, whereas pretrained networks retain an advantage in higher visual regions. This pattern generalizes to individuals excluded from model selection. Our findings establish neural-guided computational search as a route to interpretable models of vision and delineate the predictive reach of explicit, fixed computations across the visual hierarchy.