Zhengbo Jiao, Shaobo Wang, Zifan Zhang, Xuan Ren, Wei Wang, Bing Zhao, Hu Wei, Linfeng Zhang
Agentic Proposing is a framework that improves reasoning in large language models by dynamically composing modular skills, achieving high accuracy with fewer data points.
Large language models need high-quality data to improve their reasoning abilities, but creating these datasets is expensive and hard to scale. The Agentic Proposing framework addresses this by using a specialized agent that combines different reasoning skills to generate complex and verifiable problem-solving data. This method allows models to be trained with fewer, but high-quality, synthetic examples, leading to better performance even compared to large proprietary models. For instance, a model trained with this approach achieved high accuracy on a challenging math competition dataset, AIME25, showing that quality can trump quantity in data synthesis.
