Title: DIVE: Diversified Iterative Self-Improvement Show Chinese title

Title: DIVE：多样化迭代自我改进

Abstract: Recent advances in large language models (LLMs) have demonstrated the effectiveness of Iterative Self-Improvement (ISI) techniques. However, continuous training on self-generated data leads to reduced output diversity, a limitation particularly critical in reasoning tasks where diverse solution paths are essential. We present DIVE (Diversified Iterative Self-Improvement), a novel framework that addresses this challenge through two key components: Sample Pool Expansion for broader solution exploration, and Data Selection for balancing diversity and quality in preference pairs. Experiments on MATH and GSM8k datasets show that DIVE achieves a 10% to 45% relative increase in output diversity metrics while maintaining performance quality compared to vanilla ISI. Our ablation studies confirm both components' significance in achieving these improvements. Code is available at https://github.com/qinyiwei/DIVE. Show Chinese abstract

Abstract: 最近大型语言模型（LLMs）的进展展示了迭代自我改进（ISI）技术的有效性。 然而，在自生成数据上进行持续训练会导致输出多样性降低，这一限制在需要多样化解决路径的推理任务中尤其关键。 我们提出了DIVE（多样化迭代自我改进），这是一种新的框架，通过两个关键组件来解决这一挑战： 样本池扩展以实现更广泛的解决方案探索，以及数据选择以在偏好对中平衡多样性和质量。 在MATH和GSM8k数据集上的实验表明，与原始ISI相比，DIVE在输出多样性指标上实现了10%到45%的相对提升，同时保持了性能质量。 我们的消融研究证实了这两个组件在实现这些改进中的重要性。 代码可在https://github.com/qinyiwei/DIVE获取。