LLM Orchestration Causes Universal Defect Detection Cliff

A study on arXiv (2605.26174) reveals that when production language-model systems use orchestrated worker agents to answer requests, they universally lose the ability to detect cross-section defects—contradictions between distant document sections. Testing ten models across five generations from one developer and five providers with distinct alignment paradigms, researchers found a universal detection cliff: every model that could find these defects under a single agent failed under orchestration, with detection dropping two-thirds or more. The cliff is mechanism-derived and not mitigated by scale or extended reasoning. Among the six models discriminating above chance after the fall, a signal-detection decomposition showed varied behavior.