LLMZero: Discovering Adaptive Training Strategies for RL Post-Training via LLM Agents

LLMZero introduces a novel approach for discovering adaptive training strategies in reinforcement learning (RL) post-training, utilizing LLM agents to optimize training trajectories through tree search. The system identifies key distinctions in the behavior of capacity and regularization parameters, leading to significant performance improvements across four diverse GRPO tasks, achieving relative gains of 9% to 140% over the base model and 6% to 15% over grid search. This research provides valuable insights into multi-stage training design, highlighting the importance of dynamic parameter adjustments to effectively navigate exploration-exploitation trade-offs.