Searching over DARE-TIES merge coefficients of few-shot GRPO-derived LoRA directions with CMA-ES beats single-stage GRPO+LoRA on math reasoning by about 0.6 to 0.9 points while using about 10% fewer gradient updates.
arXiv preprint arXiv:2601.21598 , year=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Hyper-ES: Effective Evolution Strategies for LLM Reasoning via Descent Direction Merging
Searching over DARE-TIES merge coefficients of few-shot GRPO-derived LoRA directions with CMA-ES beats single-stage GRPO+LoRA on math reasoning by about 0.6 to 0.9 points while using about 10% fewer gradient updates.