AI "Dreaming" Mechanism Boosts Cross-Domain Discovery and Creativity

Neuroscience suggests that dreams, which often splice together unrelated elements, are not random noise but a crucial mechanism for insight and creative discovery. This research applies this concept to artificial intelligence, reframing memory consolidation as an active process of recombining knowledge across different experiences, rather than just reinforcing existing memories. The study implemented this "recombinatory-replay" mechanism in two distinct AI architectures: a LoRA fine-tuning pipeline called DREAMS and a symbolic engine named SAPIENCE. Both systems consistently showed that cross-domain consolidation significantly enhances value, while within-domain rehearsal alone does not. Specifically, the symbolic system achieved an 85.7% rate of novel cross-domain connections, a 21 percentage point improvement. The neural system saw overall gains of 5.64 percentage points, with improvements reaching 14.5 percentage points on tasks explicitly requiring cross-domain transfer, such as advanced math reasoning. This effect is inherent to the model's weights and not a prompt artifact, suggesting a fundamental principle for AI discovery.