At the 32nd National Soybean Research and Production Seminar held in Nanning, Guangxi, Professor Wang Xiaobo from Anhui Agricultural University represented the team to release the vertical large model "Fengshu" 2.0 in the soybean field. The model achieves multimodal data fusion, reduces bias from single models, and helps promote the application of soybean data in breeding.

Multi-model collaboration to reduce the bias of a single model

"Fengshu" 2.0 integrates soybean germplasm resources, genome, proteome, transcriptome, disease samples, breeding trial records, and domestic and foreign literature to form a structured knowledge system for soybean research. Compared with the first version, 2.0 introduces a multi-model collaborative analysis mechanism and integrates some general large model interfaces, allowing different models to answer the same professional question separately.

The subsequent summary module comprehensively analyzes the consensus, differences, and evidence completeness of each model's answers by combining domain knowledge, private data, and structured comparison rules, forming a clearer and more reliable professional conclusion. This design significantly improves the stability of professional knowledge services for soybeans, avoiding misleading effects caused by a single model's "hallucinations."

Six modules build the knowledge foundation for breeding

Currently, "Fengshu" 2.0 can assist in disease diagnosis, parent selection, virtual cross design, phenotypic data analysis, molecular marker-assisted breeding, and candidate gene function analysis around important agronomic traits of soybeans. It integrates six functional modules: "Soy Encyclopedia," "Soy Molecule," "Soy Literature," "Soy Disease," "Soy Phenotype," and "Soy Breeding," covering a multidimensional data system from knowledge graphs to breeding practices.

The model's underlying structure integrates over 10 million words of specialized text, more than 10,000 scientific papers and patents, building a domain knowledge graph containing 20,000 entities and 100,000 relationships. At the same time, it integrates 8,000 genome re-sequencing data, about 40,000 germplasm resources, and more than 3,000 variety phenotypic data, providing a solid data foundation for intelligent soybean breeding.