Johns Hopkins study finds LLMs discover more when trained on one-off facts

Johns Hopkins researchers found that large language models generate more novel discoveries when trained on unique, one-off facts rather than repeated information. The study suggests current data curation strategies that prioritize frequency may limit AI's scientific breakthrough potential. It recommends rethinking how training corpora are assembled, emphasizing rare facts over duplicated content. The findings matter for AI developers and data providers as model scaling increasingly depends on differentiated datasets. What to watch next is whether labs adjust data filtering pipelines to preserve singleton facts and how that affects downstream discovery benchmarks.