The integration of Large Language Models (LLM) into automatic speech recognition (ASR) system has brought remarkable improvements, making speech a fundamental interaction modality in various communication scenarios. However, challenges remain in overcoming the difficulty of huge variability in native speakers’ pronunciation and the lack of high-quality annotated dialect data, which hinder the performance of dialect recognition. Therefore, we propose a dynamic advanced reinforcement learning on Group Relative Policy Optimization (GRPO) for robust low-resource dialect ASR task. Moreover, the proposed strategy further employing curriculum learning from easy-to-hard during reinforcement learning phase significantly outperforms supervised fine-tuning (SFT) with fully randomized training data. Ablation experiments conducted on 50-hour corpus per dialect indicating the effectiveness of the proposed method for stable RL training.