HyperAIHyperAI

Command Palette

Search for a command to run...

Shanghai AI Lab launches Sage SciPerception for self-evolving scientific sensing

Recently, the Shanghai AI Laboratory introduced Sage SciPerception, a scientific multimodal framework designed to connect specialized research data with large language models. As part of the lab’s SAGE general-and-specialist architecture, the system converts scientific inputs such as spectra, biological sequences, crystal structures and time-series signals into representations that a language model can process alongside natural-language questions. By freezing the base model and training lightweight projection modules, Sage SciPerception preserves general language capabilities while injecting domain-specific scientific features into the reasoning chain. In one non-coding RNA readout experiment, only about 0.0005 billion parameters were updated, demonstrating that scientific perception can be added with minimal parameter overhead. In a 12-lead ECG example, token input was reduced from roughly 253,000 to 452, highlighting the framework’s efficiency in handling high-dimensional observational data. The framework also introduces Recursive Self-Improvement, a mode that allows the model to participate in refining its own perception architecture. Instead of relying solely on manual design and task-specific tuning, the system can analyze performance, identify weaknesses in the mapping between scientific features and language representations, propose new projection strategies, validate them through training and evaluation, and repeat the optimization cycle. In a non-coding RNA family classification task, this process raised test accuracy from 90.04 percent to 96.73 percent by revising assumptions about the performance bottleneck and introducing staged readout, encoder adaptation and independent projection pathways. On the BEACON ncRNA benchmark, accuracy reached 97.35 percent. Evaluations across ten scientific modality tasks show competitive performance in structure recognition, signal classification, site localization and property prediction. Supported by Sage SciPerception, the scientific multimodal model S2 improved capabilities in protein key-site localization, crystal property prediction and multispectral molecular identification. In MPBind protein-interaction site recognition, it achieved 61.12 percent average per-protein F1, compared with 55.53 percent for a reproduced S1-Omni. In JARVIS-DFT crystal formation-energy prediction, it reported a MAD/MAE score of 15.03. In QM9S tri-spectra molecular structure identification, its first-hit accuracy reached 49.75 percent, higher than the reported scores for S1-Omni and DiffSpectra. The system also supports traceable reasoning. For molecular spectra, it can apply chain-of-thought steps to identify structural clues, reconcile atom-count constraints and output candidate structures, allowing researchers to examine the logic behind each prediction. Looking ahead, the Shanghai AI Laboratory plans to integrate Sage SciPerception into next-generation Sage-S scientific models, expand support for new data types and tools, and continue developing self-improving scientific perception to support broader research discovery and original scientific innovation.

Related Links