Hallucination Snowball Effect: Errors Propagate in Multi-Agent LLM Pipelines

Multi-agent Large Language Model (LLM) pipelines, which chain specialized agents without immediate verification, are susceptible to a critical flaw: the "hallucination snowball effect." This phenomenon describes how errors introduced early in a pipeline don't just persist but evolve, transforming from raw facts into derived computations, then narrative prose, and finally into seemingly approved conclusions. With each transformation, the detectability of the hallucination diminishes almost irreversibly. The study formalizes this as a first-order Markov process, identifying four states of hallucination transformation: Raw Fact, Derived, Narrative, and Invisible. Empirical measurements show high escape probabilities at each stage transition, meaning errors become progressively harder to catch. In a 4-agent financial analysis pipeline, GPT-4o's detection rate dropped from 72.0% at Stage 1 to 50.9% at Stage 4, with 23.7% of hallucinations going completely undetected. Crucially, implementing verification gates *between* stages reduced hallucination survival from 58.4% to 16.2%, vastly outperforming verification only at the end of the pipeline. The research emphasizes that *when* verification occurs is more critical than *if* it occurs, recommending investment in early-stage verification.