A genuine "AI Scientist" cannot merely summarize metadata; it must deeply parse full-text technical nuances, execute multi-hop verification, and strictly demarcate scientific boundaries.
SciResearchBench provides a foundational diagnostic benchmark to bridge the gap between web browsing agents and autonomous scientific discovery.
Dataset, evaluation suite, and the DeepXiv interactive environment are fully open-sourced:
https://arxiv.org/abs/2604.25256