Reproduce “C-MOP: Integrating Momentum and Boundary-Aware Clustering for Enhanced Prompt Evolution” on OpenResearch

About this paper

Binwei Yan, Yifei Fu, Mingjian Zhu, Hanting Chen, Mingxuan Yuan, Yunhe Wang, Hailin Hu

C-MOP, from Huawei's Noah's Ark Lab, introduces a prompt optimization framework for Large Language Models that leverages boundary-aware sample selection and momentum-guided gradient clustering. This approach led to a 1.58% average accuracy improvement over leading methods and enabled a 3B parameter model to surpass 70B domain-specific models and GPT-4 on specialized tasks.

Key insights

• Systematically sampling boundary-aware examples, such as Hard Negatives, Anchors, and Boundary Pairs, significantly enhances the precision and actionability of textual gradients for prompt refinement.
• Employing a momentum-guided approach combined with cross-iteration semantic clustering effectively stabilizes prompt optimization by resolving conflicting gradients and distilling consistent improvement signals.
• The reasoning capabilities of the optimizer model are crucial; a more capable optimizer generates higher-quality textual gradients, leading to more substantial performance gains in the target task.

Summary by alphaXiv.