The Hallucination Snowball: Modeling Error Propagation as State Transitions in Multi-Agent LLM Pipelines
06:00 · August 18, 2026 · arXiv cs.AI RSS
Sequential multi-agent LLM pipelines chain specialized agents without verification at handoffs, creating a structural flaw with measurable and severe consequences. We show that hallucinations injected at Stage 1 do not merely persist; they transform: raw numerical facts become derived computations, then narrative prose, then editorially approved conclusions. At each transformation, detectability degrades near-irreversibly. We formalize this as the hallucination snowball effect, a first-order Markov process over four states (Raw Fact $\to$ Derived $\to$ Narrative $\to$ Invisible) with empirically measured per-boundary escape probabilities of 24.6%, 48.3%, and 89.3%. Across 346 automatically injected hallucinations in a 4-agent financial analysis pipeline on FinanceBench, gpt-4o detection drops from 72.0% at Stage 1 to 50.9% at Stage 4, and 23.7% of hallucinations survive completely undetected in the final output. Even the strongest model tested (Qwen3.5-397B-A17B, 87.0% at Stage 1) faces a structural ceiling; projected Stage 4 detection is only ${\sim}$60--65%. Critically, boundary gates using identical RAG verification tools reduce hallucination survival from 58.4% to 16.2% versus end-of-pipeline checking (Cohen's $h = -0.911$, $p < 0.000001$), while end-checking alone achieves merely 2.3 pp improvement over no verification. When you verify matters more than whether you verify. Our model predicts survival for $n$-agent linear pipelines and prescribes optimal verification resource allocation: invest at $S_1{\to}S_2$ first, where 75.4% of hallucinations are still catchable, not at $S_3{\to}S_4$ where 89.3% have already escaped.
Summary
Sequential multi-agent LLM pipelines, which chain specialized agents such as a researcher, analyst, writer, and reviewer without intermediate verification, allow hallucinations introduced early to transform across stages. A fabricated numerical claim can evolve into derived computations, then narrative prose, and finally an approved conclusion whose factual basis is no longer recoverable. The paper formalizes this progression as the hallucination snowball effect, represented by a first-order Markov process with four states—Raw Fact, Derived, Narrative, and Invisible—whose empirically measured escape probabilities at each boundary are 24.6 percent, 48.3 percent, and 89.3 percent.
Experiments on FinanceBench using 346 automatically injected hallucinations in a four-agent pipeline built with LangGraph and gpt-4o show that detection rates decline from 72.0 percent at the first stage to 50.9 percent at the fourth, leaving 23.7 percent of hallucinations undetected in the final output. Even stronger models exhibit the same structural decay. Boundary verification gates placed at handoff points and implemented with deterministic checks reduce hallucination survival from 58.4 percent to 16.2 percent, a 42.2 percentage-point improvement over end-of-pipeline checking alone, which yields only a 2.3 percentage-point gain relative to no verification.
The results indicate that verification timing outweighs detector strength: the majority of hallucinations remain catchable after the first transition, whereas later stages render most errors structurally invisible. The authors supply code and statistical outputs to support reproducibility and extend the model to predict survival rates for pipelines of arbitrary length.
Why it matters
Directly actionable for Dutch AI teams building reliable multi-agent systems; aligns with EU emphasis on trustworthy AI; offers novel Markov modeling and verification timing insights with high technical depth and reproducibility.
