Reproduce “LLMZero: Discovering Adaptive Training Strategies for RL Post-Training via LLM Agents” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Haoyang Fang, Wei Zhu, Boran Han, Alex Zhang, Zhenyu Pan, Shuo Yang, Shuai Zhang, Jiading Gai and 6 more

LLMZero adapts reinforcement-learning post-training by inspecting training signals and using tree search to propose checkpoint-linked phases with coordinated hyperparameter changes. In four GRPO tasks, its Qwen3-4B SSMR-Bench result was 82.2% average test accuracy across four subtasks, versus 77.6% for grid search within 16 iterations. This evidence does not isolate individual changes or establish a universal training rule.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.