Beyond AI as Assistants: Toward Autonomous Discovery in Cosmology

Researchers demonstrate two AI agent systems—CMBEvolve and CosmoEvolve—capable of autonomous scientific discovery in cosmology, moving beyond AI-as-tool toward AI-as-researcher. CMBEvolve uses code evolution for quantitative tasks while CosmoEvolve manages open-ended research workflows, both showing promising results in detecting anomalies and analyzing astronomical data without human intervention.

This research represents a meaningful shift in AI capability: autonomous systems that can formulate hypotheses, execute experiments, and iterate solutions with minimal human guidance. Rather than AI serving as a calculator or search engine, these agents operate more like junior researchers, autonomously navigating complex scientific workflows. The dual-system approach is instructive—CMBEvolve handles well-defined optimization problems where success metrics are explicit, while CosmoEvolve tackles messier, exploratory science where objectives emerge during investigation.

The cosmology domain provides an ideal testing ground because it combines both controlled benchmarks and genuinely open research questions. Weak-lensing analysis and ACT DR6 data processing are real scientific bottlenecks where human researchers spend months on routine pattern detection and diagnostic work. AI agents automating these tasks frees researchers for higher-order interpretation and theory development.

For the AI research community, this demonstrates that large language models coupled with code generation and tree-search algorithms can tackle scientific problems beyond data classification or retrieval. The systems' ability to identify "non-trivial pair- and scale-dependent behaviour" autonomously suggests these agents grasp statistical reasoning, not merely pattern matching.

The broader implications extend to reproducibility and discovery velocity. If AI agents can independently validate findings and explore parameter spaces exhaustively, scientific workflows accelerate while documentation improves automatically. Future iterations might enable multi-disciplinary agent collaboration, where systems trained on physics, data science, and domain expertise jointly solve complex problems.

• →AI agents now perform autonomous scientific discovery in cosmology, moving beyond assistant roles to independent researcher functionality.
• →CMBEvolve improves quantitative benchmarks through iterative code evolution; CosmoEvolve manages open-ended research workflows with minimal human direction.
• →Both systems produced analysis-grade results on real astronomical data, demonstrating practical value beyond theoretical proof-of-concept.
• →Cosmology provides ideal test cases combining controlled benchmarks and genuinely open problems for validating AI scientist systems.
• →Autonomous AI researchers could accelerate discovery cycles and improve reproducibility by automating routine analysis and exhaustive parameter exploration.