This paper proposes two novel strategies for improving the complex problem-solving ability of large-scale language models (LLMs): Adaptive Difficulty Curriculum Learning (ADCL) and Expert-Guided Self-Reconfiguration (EGSR). ADCL is a curriculum learning strategy that addresses the phenomenon of difficulty variation during training by re-evaluating the difficulty as the model's capabilities change. EGSR is a reinforcement learning strategy that encourages expert solutions to be restructured within the model's conceptual framework, bridging the gap between imitation learning and pure exploration, thereby promoting deeper understanding and knowledge absorption. Experimental results based on the Qwen2.5-7B model demonstrate that the two proposed strategies synergistically improve performance, achieving a 10% performance improvement on the AIME24 benchmark and a 16.6% performance improvement on the AIME25 benchmark.
