A team of 22 researchers has developed S1-Omni, a unified multimodal reasoning model that consolidates fragmented AI capabilities across scientific domains into a single system.

The model addresses a key limitation in AI for Science (AI4S), where capabilities remain scattered across domain-specific models, tool-augmented large language models, and scientific language models. This fragmentation has limited joint modeling of heterogeneous data, scientific laws, and expert knowledge.

S1-Omni's architecture rests on three core components: unified representation of scientific data, natural-world knowledge alignment, and decoding for domain-specific tasks. The system maps natural-language instructions and scientific objects — including CIF, SMILES, protein sequences, spectra, and scientific images — into a shared representation space.

The model incorporates scientific laws and expert knowledge into data construction and training, enabling reasoning from scientific evidence. It performs task-specific decoding to support property prediction, spectrum-to-molecular generation, protein site and structure prediction, and scientific image generation and editing.

Training and Performance

S1-Omni was trained on S1-Omni-Corpus, which covers 200 scientific tasks and contains millions of reasoning samples. The researchers evaluated the model on over 60 scientific benchmarks.

The model outperformed GPT-5.5 and Gemini-3.1-Pro on most benchmarks and matched or surpassed domain-specific models on several tests. This performance demonstrates the viability of unified scientific modeling without sacrificing specialized capabilities.

The research represents a significant step toward consolidating AI capabilities across scientific disciplines. Rather than requiring separate models for chemistry, biology, physics, and materials science, S1-Omni offers a single system that can handle diverse scientific reasoning tasks.

The paper was submitted to arXiv on July 17, 2026, by researchers led by Jiahao Zhao. The work provides what the authors describe as "a practical path toward unified scientific modeling" that could streamline AI applications across research domains.