Curriculum Reinforcement Learning Can Incentivize Reasoning Capacity in LLMs Beyond the Base Model

The paper introduces a boundary-aware Curriculum Reinforcement Learning (RL) approach designed to enhance the reasoning capacity of large language models (LLMs) beyond their base capabilities. By utilizing pass@k sampling to identify reasoning capacity boundaries and applying targeted teacher guidance, this method achieves significant improvements in performance metrics, with an average pass@256 score increase of 9.8 percentage points over base models like Qwen, Llama, and DeepSeek, and 10.3 percentage points over conventional Vanilla RLVR. This advancement is crucial for practitioners seeking to develop LLMs that can continually extend their reasoning abilities.