Hyper-ES Uses Descent Direction Merging for LLM Reasoning

Hyper-ES is a subspace-based evolution strategy that improves LLM reasoning by merging descent directions from inexpensive gradient-based runs. This avoids the high-dimensional instability of applying random perturbations directly to billion-parameter models.