SWE-Prime makes 10% of coding data beat all

SWE-Prime reports that training on a selected 10% of successful coding-agent trajectories beat training on the full resolved dataset, with relative gains up to 12.2% on SWE-Bench Pro and 24.2% on SWE-Bench Verified in the paper’s experiments. The selection method filters trajectories for process quality, result quality, and representativeness rather than treating every successful trace as equally useful; the released study artifacts make the claim inspectable, but it is not a promise of the same lift in production. Builders should audit for redundant or risky traces before buying more rollouts, extending the case for reusable agent skills and retrieval as a training lever.