AI for Science

AI for Science (sometimes AI4S) is the application of AI methods, especially large language models and other foundation models, to accelerate or automate parts of the scientific research process: literature review, hypothesis generation, experiment design, data analysis, simulation, and paper writing. Through the 2010s and early 2020s, AI for Science mostly meant machine learning applied to specific scientific subtasks, such as protein structure prediction or materials property prediction, with AI treated as a tool assisting human-directed research.

Starting around 2024, a further step emerged: agentic or autonomous scientific discovery, where an LLM-based system runs much of the discovery loop with limited human direction. Sakana AI's "The AI Scientist" (Lu et al., 2024, arXiv:2408.06292) demonstrated an end-to-end pipeline that generates a research idea, writes and runs code, produces figures, and writes up a full paper, at a cost of under $15 per paper. An automated reviewer scored the output above a machine learning conference's acceptance threshold. A 2025 successor, AI Scientist-v2 (arXiv:2504.08066), used agentic tree search to remove the need for a human-written starting template, and produced a manuscript that passed peer review at a real ICLR workshop.

Several 2025 surveys describe this as a shift from AI for Science to "Agentic Science," organizing autonomous systems around a small set of core capabilities: problem and hypothesis formulation, experiment design and execution, and evidence-grounded revision. Systems in this space vary widely in autonomy and reliability. Commentators note that automatically generated ideas and papers still typically need expert vetting. Open challenges include reproducibility, distinguishing genuinely novel findings from plausible-sounding ones, and determining how much human oversight autonomous research systems require.