MMKG-RDS: Reasoning Data Synthesis via Deep Mining of Multimodal Knowledge Graphs

Researchers introduce MMKG-RDS, a framework that uses multimodal knowledge graphs to synthesize high-quality training data for improving AI model reasoning abilities. Testing on Qwen3 models showed 9.2% improvement in reasoning accuracy, with applications for complex benchmark construction involving tables and formulas.

• →MMKG-RDS framework addresses limitations in existing reasoning data synthesis methods through multimodal knowledge graphs.
• →Fine-tuning Qwen3 models with synthesized data improved reasoning accuracy by 9.2% across different model sizes.
• →The framework supports fine-grained knowledge extraction and customizable path sampling for data generation.
• →MMKG-RDS-Bench dataset covers five domains, 17 task types, and 14,950 samples for validation.
• →The system generates challenging data for complex benchmarks involving tables and mathematical formulas.