Training constrains neural routes to knowledge assembly

Abstract

A hallmark of human intelligence is the ability to rapidly restructure existing knowledge when new information reveals unexpected connections, a capacity termed knowledge assembly. This cognitive flexibility distinguishes human learning from artificial systems, which catastrophically forget when acquiring new relationships. Understanding the mechanisms underlying flexible knowledge reorganization thus has implications for continual learning in both humans and algorithms. Here, using electroencephalography we show that successful knowledge assembly depends on temporally orchestrated reactivation of prior neural representations. Critically, training schedules bias learners toward different representational strategies: blocked training promotes compressed certainty-weighted codes, while interleaved training yields high-dimensional factorized representations. Vanilla recurrent networks failed to develop human-like certainty geometries despite identical training, revealing missing computational principles in current artificial systems. These findings demonstrate that cognitive flexibility emerges through creative reuse of learned representations, with training history constraining available neural routes to reorganization.