Zhuojun Ying¹ (), Marcelo Mattar², Anastasia Kiyonaga¹; ¹UCSD, ²NYU

Multi-step planning is assumed to rely on working memory (WM), but the link between the functions is poorly formalized. In recent work, we developed an information-theoretic model to examine choice behavior and WM for reward information in a sequential planning context. In behavioral testing, participants were asked to choose the most valuable path in a decision tree where nodes varied in value and were revealed sequentially. After choosing, they were asked to recall a subset of node values. Both model and behavior showed better memory for nodes that were more informative to the path choice, suggesting that limited WM is dynamically allocated during planning based on choice-relevance. The model specifies that this resource-rational allocation arises from WM constraints, and predicts that higher WM demands would shift planning toward more gist-level strategies. Here, we manipulated WM demands in two sequential decision-making experiments to test this idea. In Experiment 1, we continuously varied node color to represent values, and we manipulated WM demands by varying tree size and color distinctiveness. Recall was biased toward local (path-level) averages when node colors were more distinct, but toward more global (branch-level) averages when colors were less distinct, suggesting gist-level encoding under higher visual demands. Path choice also favored an all-or-none satisficing strategy in the latter case, consistent with the gist-level reward encoding. In Experiment 2, multiple features contributed to path quality: both color and angle varied across nodes, but only one feature was cued at the end of each path to determine the path value. Recall for choice-relevant features was biased toward the path average while recall for choice-irrelevant features was biased toward the branch average, suggesting more gist-level encoding of less informative content. Together, these findings show how visual WM demand and planning context interact to shape ensemble encoding and decision-making strategies.