Preprint
Aug 2026
Hyper-ES: Effective Evolution Strategies for LLM Reasoning via Descent Direction Merging
This work proposes Hyper-ES, a subspace-based ES framework that avoids the weakness of ES in full-parameter search while exploiting its strength in low-dimensional optimization, and consistently outperforms GRPO-LoRA while requiring 10% fewer space-consuming gradient updates.
Yuntian Gu, Zhi Zheng, Yun-Peng Ba et al.
· 0 citations