Tongyi Lab announced on June 18, 2026, the open-sourcing of LOGOS (Language Of Generative Objects in Science), the first multi-domain scientific generative foundation model based on a unified "scientific grammar."
The model was jointly developed by ATH-Token Foundry and the Huoqian Institute of Artificial Intelligence at Renmin University. It aims to break the fragmented status quo of "one task, one expert model" in the traditional AI for Science (AI4S) field. By encoding heterogeneous scientific objects such as proteins, small molecules, materials, and chemical reactions into a unified discrete token sequence, it achieves cross-domain knowledge integration and autoregressive generation under the native large model framework.
The core breakthrough of LOGOS lies in its innovative "scientific grammar" design and spatial interaction discretization technology. This enables the model to deeply understand complex 3D spatial interaction rules and achieve autoregressive generation without relying on scarce 3D coordinate data or specialized geometric networks, ensuring complete consistency in form and objective between pre-training and downstream tasks.
Evaluation data shows that LOGOS-1B, with only 1B parameters, consistently matches or exceeds domain-specific methods in six representative tasks, including pocket condition ligand generation, inverse synthesis prediction (Top-1 accuracy of 74.8%), pocket site identification (Top-n accuracy of 58.5% on HOLO4K dataset), and MOF material generation (a 76% increase in the proportion of new building units). In some tasks, it even outperformed NatureLM with 8×7B parameters, using only 1/56 of the parameters.
