DisCo Skills From 1,000 GitHub Repos Push GPT-5.5 Agent +134% on MLE-bench
Summary
Confirms that operational knowledge distilled from real GitHub repos is a separable, stackable capability layer that can more than double a frontier agent's benchmark performance with no retraining — shifting the primary lever from model weights toward external knowledge infrastructure.
Originally reported by paper
Read the original article →Original headline: DisCo Skills From 1,000 GitHub Repos Push GPT-5.5 Agent +134% on MLE-bench