Scientific discovery workflows often depend on structured curation from the literature. However, current agents face challenges because key evidence is scattered across long texts, dense tables, and figures, requiring reasoning across multiple evidence fragments rather than simply copying a single span.
To address this, we study scientific curation from multimodal sources and introduce Beaver, an agent harness that extracts structured information from scientific papers while preserving provenance to the supporting evidence.
Beaver combines a frontier agent with multimodal evidence tooling, task scaffolding, and artifact-grounded autoresearch. These components transform curation into a staged, auditable workflow, enabling an iterative evaluate-diagnose-revise loop, where persistent run artifacts expose stage-localized failures and guide harness updates.
Experiments show that Beaver achieves an 81.0 on the Gold-Referenced Attribute Score (GRAS), an attribute-level measure of agreement with gold curated records, outperforming frontier agents by over 23 absolute points.
Ablation studies demonstrate that task scaffolding, multimodal evidence tooling, and provenance traces each contribute meaningfully to performance, while attribute-level analysis indicates the largest gains on high-value attributes requiring cross-modal reasoning and normalization.
These results highlight that, for scientific curation from papers with multimodal evidence, harness design is a central determinant of agent performance.
Blogger's Review: The design of the Beaver agent significantly enhances the efficiency and accuracy of scientific literature curation through multimodal evidence integration and provenance tracking, opening new possibilities for research workflows. Its iterative feedback mechanism also provides valuable insights for future intelligent literature processing tools.