TLDRocket
Sign in

Information Extraction

1 summarised story about Information Extraction, each linking back to the original source. Browse all topics →

Friday, 17 July 2026

Building Agent Harnesses for Scientific Curation from Multimodal Sources

arXiv cs.CL 6 hours ago

Researchers introduced Beaver, an agent harness designed to extract structured information from scientific papers containing text, tables, and figures while tracking the source evidence supporting each extracted fact. Beaver achieved 81.0 on the Gold-Referenced Attribute Score, outperforming frontier agents by 23 points through combining multimodal evidence tools, task scaffolding, and iterative evaluation-diagnosis-revision loops. The system enables scientific curation workflows to become auditable and staged processes where design choices significantly impact how well agents handle complex reasoning across multiple evidence types.