Beyond AI as Assistants: Toward Autonomous Discovery in Cosmology
TLDR
Two agentic systems for cosmology, CMBEvolve and CosmoEvolve, enable autonomous discovery via code evolution and multi-agent research labs.
Reasoning
The paper presents novel systems with preliminary real-world demonstrations in cosmology, showing both controlled benchmarks and open-ended research. However, it lacks detailed evaluation metrics and comparisons to baselines, and the scope is limited to cosmology.
Read-first score
Read-first score 61.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 61.
Field roles
Rank sensitivity
Stability: volatile; rank range: 49.
Keyword Scores
Deep Analysis
Innovations
- CMBEvolve: an agentic system that uses LLM-guided code evolution and tree search for cosmology tasks with explicit quantitative objectives
- CosmoEvolve: a virtual multi-agent research laboratory for open-ended scientific workflows in cosmology
- Framing cosmology as a testbed providing both controlled benchmarks and realistic open-ended problems for developing AI scientist systems
Methodology
The paper introduces two complementary agentic systems: CMBEvolve, which optimizes explicit quantitative objectives via LLM-guided code evolution and tree search, and CosmoEvolve, a multi-agent virtual laboratory for open-ended research. Preliminary demonstrations apply CMBEvolve to out-of-distribution detection in weak-lensing maps and CosmoEvolve to autonomous analysis of ACT DR6 data.
Key Results
CMBEvolve iteratively improved the benchmark score for out-of-distribution detection in weak-lensing maps through code evolution. CosmoEvolve autonomously identified non-trivial pair- and scale-dependent behaviour in ACT DR6 data and produced analysis-grade diagnostics.
Limitations
- The work is presented as preliminary demonstrations, so the generalizability and robustness of the systems are not yet established.
- Evaluation is limited to specific cosmology tasks (weak-lensing out-of-distribution detection and ACT DR6 analysis), with no broader validation reported.