2026/02/06 by Ali Shiravand, Yumeya Yamamori, William Bhot +2 · 1 voice
Computer Science · Neuroscience · Social Sciences · #Champion #Embodied and Extended Cognition #Ethics and Social Impacts of AI #Explainable Artificial Intelligence (XAI) #Feature (linguistics) #Feature learning #Flexibility (engineering) #Raising (metalworking) #Representation (politics) #Selection (genetic algorithm) #Value (mathematics)
paper · doi:10.31234/osf.io/7bnkp_v1
openalex publication_date 2026/02/06 · openalex created_date 2026/02/08 · openalex updated_date 2026/07/14
On March 9, 2016, at the Four Seasons Hotel in Seoul, AlphaGo defeated world champion Lee Sedol 4-1, marking a milestone in machine intelligence. But unlike humans, AlphaGo pursued a single fixed goal. Humans continually revise their goals and preferences, raising the question of how learned value representations support such flexibility. Previous work on goal-dependent learning has focused on goal selection or maintenance, largely neglecting how the structure of value representations affects learning when goals change. Here, using two multi-goal learning paradigms combined with computational modelling, we compare two learning architectures: reweighting existing values versus relearning them. We show that participants using feature-based representations adjusted more quickly and monitored decisions more effectively by reusing learned feature values. In contrast, participants relying on a single composite value were forced to relearn after each switch. These findings show that value representation structure shapes learning efficiency and behavioural flexibility in multi-goal environments.