Qingqing Yang¹, Hsin-Hung Li¹; ¹The Ohio State University

Efficient coding theories posit that neural resources are optimized according to environmental statistics, successfully explaining stimulus-dependent variability in perception and working memory (WM). Recent studies further propose that rewards also influence stimulus encoding (Schaffner et al., 2023). However, past experiments provided conflicting evidence on whether trial-wise rewards modulate WM resources (Cho et al., 2022; Van Berg et al., 2023). Here, we test the hypothesis that the learnt association between stimulus features and rewards impact the allocation of WM resources. Across two sessions on separate days, participants learned two reward contexts through a 4AFC decision task (100 trials per day). In each trial, four oriented gratings (sampled from naturalistic distributions with more cardinal orientations) were presented on different quadrants and participants chose one of them to receive immediate reward. In the baseline context session, rewards were uniform across orientations, whereas in the reward-manipulated context session, rewards increased linearly as the orientations became more diagonal. After finishing the decision task, participants performed a delayed-estimation task (300 trials per day). In each trial, participants viewed a central grating (uniformly sampled orientations; set size=1; 4 dva in width, 500 ms duration) and reproduced its orientation after a 1s or 5s delay. The rewards in each trial were based on the memory error, and scaled by the orientation-reward mapping learned during the decision task. Across sessions, behavioral estimations showed higher variability at more diagonal orientations (45°, 135°), consistent with the oblique effect. Critically, we found an interaction between orientations and reward contexts. Orientations with higher diagonality were remembered with greater precision—lower variability and smaller error—when associated with higher reward compared to the baseline context. This effect is stronger with a longer (5s) delay. Our results indicate that participants actively adjusted the allocation of WM resources during maintenance based on the expected utility of features.