FlyAOC: Evaluating Agentic Ontology Curation of Drosophila Scientific Knowledge Bases
AI agents are now being tested to read and summarize thousands of scientific papers automatically.
Deep Dive
Researchers have launched 'FlyBench,' a new benchmark to evaluate AI agents on the complex, end-to-end task of curating scientific knowledge from literature. The test requires agents to search and read from a corpus of 16,898 full-text papers about fruit fly (Drosophila) genes to produce structured, expert-level annotations. Multi-agent AI architectures performed best, but all current models leave significant room for improvement, struggling to discover new information versus confirming what they already know.
Why It Matters
This could massively accelerate scientific discovery by automating the tedious curation of research findings into usable databases.