Competing with AI Scientists: A Fully Agent-Driven Approach to Cosmological Parameter Inference

We present one of the winning solutions for Phase 1 of the FAIR Universe Weak Lensing ML Uncertainty Challenge, achieved through a framework of autonomous AI Agents. We use CMBAgent, an LLM-driven framework using planning and control strategy to autonomously navigate the scientific landscape from idea generation and code execution to validation and refinement combined with one-shot prompting and human-in-the-loop. This agent-led workflow implemented an Inception-style CNN architecture with D4 group data augmentation. Furthermore, the agents autonomously developed a covariance estimation technique with a grid-based inference method. Complementing the agent-driven strategy, human experts incorporated Scattering Transform data into the training and inference pipeline. To the best of our knowledge, this marks the first demonstration of AI agents achieving top-tier performance in a complex, real-world cosmological data analysis task.