Hallucination Snowball: Error Propagation in Multi-Agent LLM Pipelines

An arXiv paper (2608.14588) introduces the concept of the 'hallucination snowball effect' within sequential multi-agent LLM pipelines, illustrating how early-stage hallucinations evolve into forms that are harder to detect. The authors represent this phenomenon as a first-order Markov process consisting of four states: Raw Fact, Derived, Narrative, and Invisible, with escape probabilities measured at 24.6%, 48.3%, and 89.3%. In a financial analysis pipeline involving four agents on FinanceBench, they introduced 346 hallucinations, revealing a decline in gpt-4o detection from 72.0% at Stage 1 to 50.9% by Stage 4, with 23.7% of hallucinations remaining completely undetected in the final output. The study underscores a critical flaw in multi-agent systems, where agents lack verification during transitions, leading to the transformation of numerical facts into derived computations, narrative text, and ultimately, editorial conclusions. The research offers quantitative insights into the degradation of detectability through these transformations. The paper was published on arXiv under ID 2608.14588v1.