Efficient Reinforcement Finetuning via Adaptive Curriculum Learning
This work addresses the sample- and compute-inefficiency problem in reinforcement finetuning for researchers and practitioners, offering a more scalable framework, though it is incremental as it builds on existing methods like PPO.
The paper tackles the inefficiency of reinforcement finetuning for large language models in mathematical reasoning by introducing AdaRFT, which uses adaptive curriculum learning to dynamically adjust problem difficulty, resulting in up to 2x faster training and improved accuracy on competition-level math datasets.
Reinforcement finetuning (RFT) has shown great potential for enhancing the mathematical reasoning capabilities of large language models (LLMs), but it is often sample- and compute-inefficient, requiring extensive training. In this work, we introduce AdaRFT (Adaptive Curriculum Reinforcement Finetuning), a method that significantly improves both the efficiency and final accuracy of RFT through adaptive curriculum learning. AdaRFT dynamically adjusts the difficulty of training problems based on the model's recent reward signals, ensuring that the model consistently trains on tasks that are challenging but solvable. This adaptive sampling strategy accelerates learning by maintaining an optimal difficulty range, avoiding wasted computation on problems that are too easy or too hard. AdaRFT requires only a lightweight extension to standard RFT algorithms like Proximal Policy Optimization (PPO), without modifying the reward function or model architecture. Experiments on competition-level math datasets-including AMC, AIME, and IMO-style problems-demonstrate that AdaRFT significantly improves both training efficiency and reasoning performance. We evaluate AdaRFT across multiple data distributions and model sizes, showing that it reduces training time by up to 2x and improves accuracy by a considerable margin, offering a more scalable and effective RFT framework.