Imaging-101: Benchmarking LLM Coding Agents on Scientific Computational Imaging

Imaging-101 introduces 57 expert-verified computational imaging tasks across six scientific domains. Each task is grounded in a peer-reviewed paper and standardized into four stages: preprocessing, forward physics modeling, inverse solving, and visualization. The benchmark evaluates seven frontier LLMs through planning, function-level unit tests, and end-to-end reconstruction tracks. Its reported findings identify difficulties that general coding benchmarks may miss, especially algorithm selection, physical convention handling, and integrating complete reconstruction pipelines. The authors suggest skill-augmented and domain-specialized agents as a practical direction for improving reliability.