LLM orchestration creates universal cliff for cross-section defects

A new paper from researcher Hiroki Fukui reveals a fundamental blind spot in orchestrated AI systems. When a single LLM is asked to detect contradictions between two distant sections of a document, it can spot the defect. But when the same request is split across multiple worker agents that recompose a single integrated report, detection rates collapse by at least two-thirds across every model and paradigm tested—from OpenAI to Anthropic to open-source families. The author calls this a "universal detection cliff," and it resists fixes like scaling model size or extending reasoning steps. The mechanism is baked into the orchestration itself: no individual agent sees the full context, so structural inconsistencies between partitions become invisible.

Beyond the cliff, the paper identifies a dangerous pattern in how models behave after falling off it. Using signal-detection analysis, only one developer's model generations showed a systematic criterion shift: as alignment training strengthened, the model missed fewer defects but flooded clean documents with false alarms. The most aligned systems, the author warns, are not the safest. The integrated report often signals high confidence even while missing a critical contradiction—a false sense of security. The study releases all data and scripts, and suggests that confidence scores from orchestrated reports are uninformative about partition-spanning defects. This has immediate implications for any production system that chains multiple LLM agents to analyze long documents.