Multi-stage efficient coding of perception and value in goal-directed behavior

Abstract

To act effectively, the brain must transform information through a chain of processing stages, from sensing the environment, to evaluating options, to selecting actions. Because neural resources are limited, each stage should represent information efficiently. Yet efficient coding has been studied almost exclusively in perception, and always one stage at a time. Whether perception and valuation are each governed by their own efficient code, and how these codes interact along the pathway from sensation to action, remains an open question. We developed a formal framework, comparing models in which efficient coding and Bayesian decoding shape perception only, valuation only, or both. To tease apart contributions from each stage, we designed an experiment that independently varied how stimuli map onto values. In a preregistered study, behavior was best explained by efficient coding operating at both stages, with each stage tracking its own objective prior. Crucially, changing the value distribution reversed the classic repulsion biases seen in orientation perception, revealing separable efficient codes in perception and valuation. The two stages also operated on different timescales: perceptual representations reflected stable, long-term environmental structure, while value representations updated rapidly with changing context. Together, these findings show that separable efficient codes at successive processing stages combine to shape behavior. This principle likely extends well beyond perception and valuation: whenever behavior depends on abstract, constructed representations rather than raw sensory signals, the brain may solve the efficiency problem anew at each stage.