Research map / Large language models / Instruction and fine-tuning
Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning
Xin Qiu, Yulu Gan, Conor F. Hayes, Qiyao Liang, Yinggan XU, Roberto Dailey, Elliot Meyerson, Babak Hodjat, Risto Miikkulainen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ES instead of Dr.GRPO-v | Qwen-2.5-3B-Instruct · Countdown | accuracy (%): 43.8 -> 60.5 | Improved |
| use ES instead of GRPO-v | Llama-3.1-8B-Instruct · Countdown | accuracy (%): 46.9 -> 61.2 | Improved |
| use ES instead of Dr.GRPO-v | Qwen-2.5-0.5B-Instruct · Countdown | accuracy (%): 13.5 -> 14.4 | Improved |
| use ES instead of GRPO-v over three seeds | Qwen2.5-3B-Instruct · Countdown | accuracy (%): 35.07 -> 58.67 | Improved |
| use ES instead of GRPO with KL penalty | Qwen-2.5-7B-Instruct · Conciseness | reward and KL: reward 0.911 -> 0.919; KL 1.591 -> 0.813 | Improved |
| remove KL penalty from GRPO (beta 0.0 and 0.01) (ablation) | GRPO · Conciseness | reward hacking | Hurt |
| use ES instead of GRPO for run-to-run consistency | Qwen-2.5-7B-Instruct · Conciseness | std of reward: 15.5x higher standard deviation for GRPO | Improved |
| ES fine-tune base model on ARC-AGI | Qwen-2.5-14B · ARC-AGI | accuracy (%): 0.2 -> 29.5 | Improved |
| ES fine-tune base model on Sudoku | Qwen-2.5-3B · Sudoku | accuracy (%): 2.5 -> 69.5 | Improved |
| increase ES population size from 16 to 128 (ablation) | ES · GSM8K Qwen2.5-1.5B | pass@1: 0.697 -> 0.727 | Improved |
| increase ES population size from 16 to 128 (ablation) | ES · GSM8K Qwen2.5-3B | pass@1: 0.807 -> 0.831 | Improved |
| increase ES population size from 30 to 64 (ablation) | ES · GSM8K Qwen2.5-3B | pass@1: 0.823 -> 0.804 | Hurt |
Closest papers by meaning
- ESSAM: A Novel Competitive Evolution Strategies Approach to Reinforcement Learning for Memory Efficient LLMs Fine-Tuning (NeurIPS 2026 · similarity 0.90)
- On Predictability of Reinforcement Learning Dynamics for Large Language Models (ICLR 2026 · similarity 0.89)
- ShiQ: Bringing back Bellman to LLMs (NeurIPS 2025 · similarity 0.89)
- Coevolving with the Other You: Fine-Tuning LLM with Sequential Cooperative Multi-Agent Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- UFO-RL: Uncertainty-Focused Optimization for Efficient Reinforcement Learning Data Selection (NeurIPS 2025 · similarity 0.88)
- EVOLvE: Evaluating and Optimizing LLMs For In-Context Exploration (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
