LOGOS: A Unified Generative Foundation Model for Natural Sciences

LOGOS is a scientific generative language model based on a shared scientific grammar. Its core innovation lies in representing spatial contact and constraint patterns as discrete tokens, enabling unified processing of heterogeneous tasks across natural science domains and providing feasibility evidence for the concept of 'one model fits all scientific tasks'.

Artificial intelligence has made significant progress in the field of scientific research (AI4S), but most methods face the problem of domain isolation. Different domains require specialized model architectures and training processes, which increases R&D costs and hinders cross-domain knowledge transfer. Traditional scientific computing models rely on explicit coordinates and geometric neural networks, making it difficult to scale to a wide range of scientific problems; the large differences in data representation and task objectives across domains make building a general-purpose scientific AI system extremely challenging.

LOGOS proposes a revolutionary unified framework that encodes various scientific objects and their spatial interactions into token sequences from a shared vocabulary, processing cross-domain tasks in an autoregressive manner. Its technical breakthroughs lie in the discrete representation of spatial interactions (freeing from dependence on explicit coordinates) and the unified task formulation (transforming downstream tasks into token prediction problems, achieving strong alignment between pre-training and downstream objectives).

The research team trained three LOGOS models with 1 billion, 3 billion, and 8 billion parameters. Experiments show a positive correlation between model scale and performance. In multi-domain benchmark tests, LOGOS achieved or exceeded the performance of domain-specific baseline models, initially proving the feasibility of the 'one model fits all scientific tasks' concept.

LOGOS represents an important step towards the generalization of scientific AI, pointing the way for the development of scientific computing infrastructure. Its implications are that the future of AI4S should be deeply aligned with large language models (LLMs), sharing architectures, training paradigms, and reasoning infrastructure. LLMs can become new entry points for AI4S, lowering technical barriers and promoting ecological development.

The LOGOS team has open-sourced model weights and related resources to help researchers explore the potential of unified scientific models and develop new application methods, which aligns with the trends of reproducibility and collaborative research in the AI field.