🤖 AI Summary
This study addresses the challenge large language models face in detecting cross-paragraph structural inconsistencies during multi-agent collaborative long-document generation. By fixing document content, defect types, and evaluation protocols, the authors systematically evaluate ten prominent models under both single-agent and multi-agent settings. Leveraging signal detection theory decomposition, controlled experiments, private record reconstruction, and automated scoring, they find that all models exhibit a performance drop exceeding two-thirds in multi-agent coordination scenarios. Only one developer-provided model shows a significant shift in reporting criteria (p<0.001), yet its confidence scores fail to reflect cross-segment defects. This work is the first to reveal the “detection cliff” phenomenon and introduces a reproducible benchmark framework for future research.
📝 Abstract
Production language-model systems answer a request by partitioning it across an invisible orchestration of worker agents that recompose one integrated report. We ask what this does to a class of defect no single worker can see: a contradiction in the relation between two distant sections of a document. Holding the documents, defects, mechanism, scoring, and seed fixed, we vary only the model -- ten systems across five generations from one developer and five providers from distinct alignment paradigms.
Two layers separate. First, a universal detection cliff: every model that finds these cross-section defects under a single agent loses that ability under orchestration, detection falling two-thirds or more across every paradigm tested. The cliff is mechanism-derived and not closed by scale or extended reasoning. Second, how models behave once fallen. A signal-detection decomposition shows that, among the six models discriminating above chance, only one developer's generations move along the reporting-criterion axis: as alignment is strengthened, the model misses fewer defects yet raises more false alarms on clean documents -- two faces of one criterion shift, scaling with generation within that developer (p < 0.001) and near-absent elsewhere.
At the floor the missed defect is often not out of view: the model's private record reconstructs the structural fault accurately, while the integrated report signs off on its soundness, its concern spent on the artifact and an absent collaborator. This resists quantification -- an automated judge is unstable (precision 17-50%) and keywords cannot separate it from ordinary agreement -- a resistance we report as a finding. We release all runs, probes, defect keys, scorer prompts, and scripts. An integrated report's confidence is uninformative about partition-spanning defects, the most aligned systems are not the safest, and the cliff is structural.