MMKG-RDS: Reasoning Data Synthesis via Deep Mining of Multimodal Knowledge Graphs
New method synthesizes 14,950 training samples across 5 domains, challenging models on tables and formulas.
A research team has introduced MMKG-RDS, a novel framework designed to synthesize high-quality training data specifically to enhance AI models' reasoning capabilities. The system addresses key limitations in current data synthesis methods, such as poor coverage of long-tail knowledge and lack of interpretability, by deeply mining multimodal knowledge graphs (MMKGs). These graphs combine textual, visual, and potentially other data types into interconnected networks of facts. The framework's flexibility allows for fine-grained knowledge extraction and customizable sampling of reasoning paths, enabling the generation of targeted, high-quality question-answer pairs for training. The accompanying MMKG-RDS-Bench dataset, spanning five domains and 17 distinct task types with 14,950 samples, serves as both a validation tool and a resource for the community.
The technical validation demonstrates significant impact: fine-tuning various sizes of Qwen3 models (0.6B, 8B, and 32B parameters) on a relatively small set of synthesized data from MMKG-RDS yielded an average 9.2% improvement in reasoning accuracy. Crucially, the framework excels at generating 'distinct' data that existing models find challenging, particularly for tasks involving tables and mathematical formulas. This capability is vital for constructing more rigorous and complex benchmarks to stress-test AI systems. By providing a systematic, customizable, and interpretable pipeline for creating reasoning-focused training data, MMKG-RDS offers a powerful tool for developers aiming to boost model performance in specialized domains without massive manual data collection. The code and dataset are publicly available, facilitating further research and application.
- Framework synthesizes 14,950 high-quality reasoning samples across 5 domains and 17 task types.
- Fine-tuning Qwen3 models (0.6B/8B/32B) with the data boosted reasoning accuracy by 9.2%.
- Generates challenging data involving tables and formulas, useful for building complex AI benchmarks.
Why It Matters
Provides a scalable, automated method to create targeted training data, significantly improving AI reasoning for specialized applications and benchmarks.