Large language and foundation models in predictive plant breeding: Capabilities, evidence, and challenges

Plant breeding is entering a multimodal era in which genomic, phenotypic, environmental, and multi-omics data are generated alongside a rapidly expanding scientific literature. Text-based large language models (LLMs), retrieval-augmented generation, knowledge graphs, and biological foundation models (Bio-FMs) offer complementary capabilities for organizing evidence, representing biological relationships, and learning from molecular data. Here, we review these technologies across knowledge mining, predictive modeling, and intelligent breeding design, and apply a systematic evidence audit to distinguish what each model family has directly demonstrated, where conclusions depend on adjacent or non-LLM evidence, and where key validation gaps remain. The audit shows that direct text-based LLM evidence is concentrated in retrieval, question answering, and evidence organization. Molecular prediction tasks, such as functional annotation and variant-effect prediction, rely mainly on Bio-FMs, whereas quantitative evidence for genomic prediction and G×E still comes largely from quantitative-genetic and deep-learning models. LLM-informed genomic prediction, parental selection, and breeding design therefore remain prospective. Moving them into practice will require reliable provenance, appropriate benchmarking, and validation matched to each output, from computational evaluation to experiments and field trials. LLM and Bio-FM outputs should support, rather than replace, scientific interpretation and breeder decision-making.