self-driving-labs

This paper presents a benchmark and trace-logging framework for evaluating LLM-based agents that control microscopes, comparing 105 agent configurations and finding that benchmarks support qualification but do not reliably predict performance on unseen tasks.