S1-Omni: A Unified Multimodal Reasoning Model for Scientific Understanding, Prediction, and Generation
S1-Omni is presented as one model spanning scientific understanding, prediction, and generation across many data types.
The paper says it maps instructions and scientific objects, including CIF, SMILES, protein sequences, spectra, and images, into a shared representation space. It adds scientific laws and expert knowledge during data construction and training, then uses task-specific decoding for outputs such as property prediction, molecule generation, protein prediction, and image editing. The authors trained it on S1-Omni-Corpus, covering 200 scientific tasks, and evaluated it on more than 60 benchmarks. They report that it beats GPT-5.5 and Gemini-3.1-Pro on most benchmarks and matches or exceeds some domain-specific models. HF Daily Papers' note
The paper says it maps instructions and scientific objects, including CIF, SMILES, protein sequences, spectra, and images, into a shared representation space. It adds scientific laws and expert knowledge during data construction and training, then uses task-specific decoding for outputs such as property prediction, molecule generation, protein prediction, and image editing. The authors trained it on S1-Omni-Corpus, covering 200 scientific tasks, and evaluated it on more than 60 benchmarks. They report that it beats GPT-5.5 and Gemini-3.1-Pro on most benchmarks and matches or exceeds some domain-specific models. HF Daily Papers' note
score 5