CVEvolve: AI Agent Discovers Scientific Data Algorithms Automatically
CVEvolve is an autonomous agentic harness with a zero-code interface designed to discover algorithms for scientific data processing. It targets domain scientists lacking computing or image-processing expertise, especially when data is noisy, high dynamic range, sparsely labeled, or loosely specified. The system combines multi-round search with code execution, evaluation, history management, holdout testing, and optional inspection of data and visual outputs. Its search alternates between discovery and improvement actions, using lineage-aware stochastic candidate sampling to balance exploration and exploitation. Demonstrated applications include x-ray fluorescence microscopy image registration and Bragg peak detection. The paper is available on arXiv with ID 2605.11359.
Key facts
- CVEvolve is an autonomous agentic harness for algorithm discovery.
- It features a zero-code interface for domain scientists.
- Addresses noisy, high dynamic range, sparsely labeled, or loosely specified data.
- Uses multi-round search with code execution and evaluation.
- Includes history management, holdout testing, and visual inspection.
- Search alternates between discovery and improvement actions.
- Employs lineage-aware stochastic candidate sampling.
- Demonstrated on x-ray fluorescence microscopy image registration and Bragg peak detection.
Entities
Institutions
- arXiv