Can Recombination Displace Dominant Scientific Ideas

Scientific breakthroughs are widely attributed to the broad recombination of existing knowledge. Yet despite the explosive growth of scientific labor and publications - expanding opportunities for recombination - breakthroughs have not kept pace. To investigate this disconnect, we analyze 41 million papers published between 1965 and 2024. We quantify each paper's atypicality, defined as the recombination of distant knowledge, and its disruption, which we interpret as an indicator of breakthrough innovation. Contrary to recombinant growth theory, we find a robust negative correlation between atypicality and disruption - consistent across fields, time, team size, and even versions of the same paper. Drawing on scientist interviews and large-scale bibliometric analysis, we find that atypicality reflects the extension of dominant ideas through cross-topic recombination, whereas disruption captures their replacement within the same topic - suggesting that recombination tends to consolidate prevailing paradigms, whereas disruption challenges them. Using large language models to distinguish method and theory oriented papers, we show that methods are harder to displace than theories, revealing distinct temporal dynamics in epistemic change.