MetaCurate preprint reports better coding-agent training data selection

Researchers report that MetaCurate, a framework for selecting coding-agent training examples, achieved the highest mean SWE-Verified score among the methods compared in their preprint. Its comparisons included hand-authored filtering and a baseline whose search required repeated model training and cost 3.5x as much. MetaCurate evolves executable filters using a language model's judgments of pairs of agent runs. Disagreements guide revisions to the filters, which then score runs individually. Researchers tested downstream training results on an uncurated trajectory pool and the Nemotron v3 SWE dataset; the reported advantage applies to those experiments.