Value generalization, not normalization, drives learning in structure-based grouped contexts.

Value learning is often studied at the individual object level, yet in natural environments, values are learned within structured contexts. A dominant account-value normalization-proposes that item values are encoded relative to their local context, producing contrastive biases that suppress below-average values and amplify above-average ones. However, empirical support for value normalization has been limited to proximity-based grouping (i.e., grouping by spatial or temporal co-occurrence), leaving open how value learning operates in more structured environments. Here, we focus on structure-based grouping, in which items are related by conceptual or relational structure and test value generalization as a competing account of context-dependent value learning. Across five experiments using modified multiarmed bandit tasks, participants learned item values through trial and error, while group structure had to be inferred from experience. Contrary to the predictions of value normalization, participants systematically overestimated the value of below-average items in high-mean groups and preferred them over objectively superior items from lower mean groups. This assimilative bias was robust across groupings defined by visual features, task-defined categories, and abstract cognitive map structure and persisted even when explicit monetary reward framing was removed. Computational modeling showed that behavior was best captured by value generalization mechanisms, implemented either as feedback-driven value propagation within groups or as Bayesian integration of group-level and item-level information at decision time. Across experiments, these models consistently outperformed normalization and standard item-level learning models. Together, these findings challenge the dominance of value normalization and identify value generalization as a cognitively efficient-though imperfect-strategy for value learning in structured environments. (PsycInfo Database Record (c) 2026 APA, all rights reserved).