This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
QuestA: Expanding Reasoning Capacity in LLMs via Question Augmentation
Created by
Haebom
Author
Jiazheng Li, Hongzhou Lin, Hong Lu, Kaiyue Wen, Zaiwen Yang, Jiaxuan Gao, Yi Wu, Jingzhao Zhang
Outline
Reinforcement learning (RL) has emerged as an important paradigm for training large-scale language models (LLMs) for inference tasks, but research has shown that it has limitations in encouraging the inference capabilities of existing models beyond their capabilities. This paper proposes Question Augmentation (QA) to effectively solve more challenging inference problems using RL. This method introduces partial solutions during training to reduce the difficulty of the problem and provide more informative learning signals. The proposed QuestA method was applied during RL training for a mathematical inference task and improved performance not only in pass@1 but also in pass@k. This enabled continuous improvement while further enhancing the inference capabilities of powerful open-source models such as DeepScaleR and OpenMath Nemotron. Using a 1.5B parameter model, QuestA achieved new state-of-the-art results of 72.50% (+10.73%) on AIME24, 62.29% (+12.79%) on AIME25, and 41.67% (+10.11%) on HMMT25.
Takeaways, Limitations
•
Takeaways:
◦
Improving the effectiveness of RL training through question augmentation to enhance mathematical reasoning problem-solving ability.
◦
Further improving the performance of existing powerful open source models.
◦
Achieving new SOTA with 1.5B parameter model.
•
Limitations:
◦
There is no specific mention of Limitations in the paper.
