SpecX benchmark brings 1.7M molecules to multimodal spectroscopy AI

SpecX is a new large-scale benchmark for multi-modal spectroscopy, containing 1.7 million molecules with spectral data across seven modalities: NMR (1H, 13C, HSQC), IR, MS, UV, Raman, and FL. It is organized into three tiers—a massive pretraining dataset, an aligned multi-spectral subset for benchmarking, and a high-quality experimental subset for validation. The benchmark supports tasks such as molecular elucidation (identifying molecules from spectra), spectrum simulation (predicting spectra from molecules), and spectral understanding (answering questions about spectral features). This structure enables unified evaluation across both specialized spectral models and multimodal large language models (MLLMs).

Initial experiments reveal stark performance differences: specialized models excel at signal-level modeling (e.g., precise peak identification), while MLLMs demonstrate superior high-level reasoning (e.g., interpreting spectral patterns) but lack precise spectral grounding, often hallucinating peak positions. The authors argue that current MLLMs are not ready for real spectral analysis without fine-tuning, and that the field needs spectrum-native foundation models—models trained from scratch on spectral data rather than repurposed from text or images. SpecX provides the first unified cross-paradigm benchmark to drive this development, filling a critical gap in chemistry and materials science AI.