Characterizing the Evolution Strategies Paradox in LLM Optimization

Research identifies a paradox in Evolution Strategies where offspring models perform worse than the parent, yet the parent model improves during training. This phenomenon occurs because high-dimensional exploration identifies beneficial directions that, when aggregated, allow for local improvements in high-curvature landscapes.