arXiv 'From Spark to Fire': Single Error Seed Cascades to System-Wide False Consensus in Multi-Agent LLMs — Training-Free Plugin Fixes from 32% to 89% Defense

arXiv·high signalXBlueskyLinkedInCopy link

Xie et al. demonstrate that minor inaccuracies in multi-agent LLM systems solidify into system-level false consensus through three failure modes: cascade amplification, topological sensitivity, and consensus inertia. A training-free message-layer plugin raises defense success from 32% to 89% without modifying agent architecture or retraining. Directly actionable for any pipeline with chained Claude/GPT calls where one bad retrieval or hallucination can contaminate downstream agents.