A Survey of Scientific Large Language Models: From Data Foundations to Agent Frontiers

*AI collaborators for science*

Summary This survey reviews over 270 datasets and 190 benchmarks for scientific LLMs (Sci-LLMs). It traces the evolution of models from data-focused pretraining to agentic frontiers where LLMs autonomously propose, test, and refine hypotheses

💡 Intuition Science is not just reading papers; it’s about experimenting and validating. Sci-LLMs should be more like lab partners than search engines.

🎯 Problem General LLMs are poorly equipped for multimodal, uncertainty-rich, and domain-specific data, and existing evaluations don’t capture scientific workflows.

🛠️ Solution

The survey organizes Sci-LLMs into:

• Data-Centric: pretrained on large, domain corpora.
• Domain-Specific Models: tailored for fields like chemistry or physics.
• Agentic Sci-LLMs: interactive models that design experiments and run analyses. It advocates for closed-loop ecosystems where LLMs don’t just process knowledge but also actively contribute to discovery. The paper also emphasizes reproducibility, open data, and benchmark standardization to fairly assess progress across disciplines.

🚧 Limitations and Future Opportunities

• Data scarcity and biases remain.
• Future: multimodal integration (text, images, simulation data) and autonomous “lab-bot” systems.