A benchmark of 169 expert-curated molecular dynamics simulation tasks that exposes fundamental limitations of current AI agents in scientific computing — the best agent solves only 21% of easy-level tasks.

A position paper identifying four fundamental limitations of current agentic AI scientists — problem selection bias, LLM training gaps in laboratory practice, loss of output diversity, and benchmarks that lack physical feedback loops — that prevent true autonomous scientific discovery.

A framework that uses calibrated AI model probabilities to identify cost-effective human-AI collaboration strategies for classification tasks, evaluated on CIFAR-10H and ImageNet-16H.