Adaptive-GEPA extends the GEPA reflective optimizer to handle heterogeneous requests. It evolves a router and a library of specialist programs, where each specialist pairs a description with executable code. The router decides which specialist handles each request. All components are editable text, changed via feedback from execution traces.
On a fixed mixture of four task families using Qwen3-8B, Adaptive-GEPA learned four specialists without any task labels. Its routing matched the true task partition on all 651 test requests. The family-mean test score rose from 52.6 to 70.6, outperforming GEPA's full-program adapter (62.5) and GRPO (54.0) at a nominal budget of 18,000 scored calls.
This is interesting because it learns both how to divide work and how to solve each part, adapting to heterogeneous requests without requiring pre-specified categories or separate optimization per family.