Title:CDS: Data Synthesis Method Guided by Cognitive Diagnosis Theory

Abstract:Large Language Models (LLMs) have demonstrated outstanding capabilities across various domains, but the increasing complexity of new challenges demands enhanced performance and adaptability. Traditional benchmarks, although comprehensive, often lack the granularity needed for detailed capability analysis. This study introduces the Cognitive Diagnostic Synthesis (CDS) method, which employs Cognitive Diagnosis Theory (CDT) for precise evaluation and targeted enhancement of LLMs. By decomposing complex tasks into discrete knowledge points, CDS accurately identifies and synthesizes data targeting model weaknesses, thereby enhancing the model's performance. This framework proposes a comprehensive pipeline driven by knowledge point evaluation, synthesis, data augmentation, and filtering, which significantly improves the model's mathematical and coding capabilities, achieving up to an 11.12% improvement in optimal scenarios.