Dreaming AI beats standard models: cross-domain recombination yields +14.5% math reasoning

A new paper from Zahn, Evans, and Eagleman titled "Discovery by Dreaming" proposes that AI systems need a separate offline consolidation phase—analogous to human dreaming—to recombine knowledge across domains and generate genuine insights, rather than just recalling training data. The team tested this by implementing two architecturally unrelated systems: DREAMS (a LoRA fine-tuning pipeline) and SAPIENCE (a symbolic engine that replays structured knowledge objects). Both converged on the same result: cross-domain consolidation creates measurable discovery value, while within-domain rehearsal does not.

Specifically, the symbolic arm (SAPIENCE) surfaced 85.7% novel cross-domain connections—a +21 percentage point improvement over baseline. The neural arm (DREAMS) improved overall by +5.64 pp, but on subtasks explicitly requiring cross-domain transfer (e.g., unseen math reasoning on GSM8K), gains reached +14.5 pp. The researchers confirmed the effect is a genuine property of the model weights by showing that prepending the same cross-domain material in-context to a massive 671B-parameter model actually reversed the gain. They validated their predictions against 50,000 real scientific papers and state a falsifiable hippocampal-recording prediction. The core takeaway: consolidation is not for remembering, but for discovering.