GEPA Optimization Loop Gets 3-4x Faster With Parallel Candidate Evaluation

dbreunig · x · 2026-08-06

The optimization loop for GEPA has been significantly improved for speed and generalization. Instead of proposing and evaluating a single candidate per step, the new update introduces batch proposals evaluated concurrently.

This parallelization cuts wall-clock time by 3-4x. Surprisingly, the batched approach also yields better generalization, improving performance by up to 11 points.

Original post →

More from Research

Research channel →