Trained RNNs suggest possible mechanisms. First, asymmetric feedback from decision to memory populations push memory states toward the chosen directions. Second, a warped representational geometry of orientations induces drift, amplifying decision-consistent bias near diverging stimuli. (11/13)

Next, going beyond sparse behavioral measurements, we tracked working memory states using simultaneously recorded BOLD signals from early visual cortex. Combined decoding and event-based analyses reveal neural signatures consistent with our drift-diffusion scenario. (9/13)

The model makes nuanced predictions about how bias evolves after choice. First, earlier decisions during a delay allow more time for decision-consistent biases to grow. Second, this growth is pronounced near diverging points, and reduced around attractors—consistent with human behavior. (8/13)

We tested our model predictions with a paradigm where participants performed both discrimination and estimation tasks on the same stimulus. Long delays between tasks allowed us to track slow memory dynamics through both behavior and fMRI. (5/13)

We modeled these scenarios using a drift-diffusion framework constrained by efficient coding. Drift shifts memory in a stimulus-specific way and can amplify the decision-consistent bias by biasing the choice and then continuing to bias memory in line with that choice. (4/13)