Title: Moving Past AI Assistance: The Era of Autonomous Discovery in Cosmology

Abstract: The latest developments in artificial intelligence (AI) agents are shifting the paradigm from AI as mere tools to AI capable of autonomous scientific discovery. This paper explores two complementary agentic frameworks designed for cosmology: \texttt{CMBEvolve} and \texttt{CosmoEvolve}. The former addresses tasks with clear quantitative goals by utilizing LLM-guided code evolution and tree search, while the latter facilitates open-ended scientific workflows via a virtual multi-agent research laboratory. In preliminary trials, \texttt{CMBEvolve} was applied to out-of-distribution detection within weak-lensing maps, where it progressively enhanced benchmark scores through iterative code evolution. Meanwhile, \texttt{CosmoEvolve} was deployed for the autonomous analysis of ACT DR6 data, successfully uncovering non-trivial pair- and scale-dependent behaviors and generating analysis-grade diagnostics. These case studies illustrate how cosmology offers both controlled benchmark tasks and complex, open-ended research challenges ideal for advancing AI scientist systems.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC