EMNLP paper traces citation errors in deep research systems: the orchestrator causes most

EliasEskin · x · 2026-09-23

An EMNLP 2026 main-conference paper diagnoses citation errors in deep research systems, where many sentences lack supporting citations — recall is 58.7% in NVIDIA AI-Q and just 7.1% in TrajectoryKit. Information gets corrupted between agents like a game of telephone; the authors propose an algorithm that traces each error to the responsible agent and classifies its type. Surprisingly, the orchestrator causes most final-report errors despite having one of the lowest per-agent error rates, and a one-sentence prompt fix raises citation recall by 5%.

Related event: EMNLP Paper Finds Deep Research Reports' Citation Support as Low as 7.1%(2 posts)→

Original post →

More from coding & agent

coding & agent channel →