The Hallucination Snowball - Modeling Error Propagation as State Transitions in Multi-Agent LLM Pipelines

Sep 4, 2026

Sequential multi-agent LLM pipelines chain specialized agents at handoffs without verification, causing hallucinations to propagate through successive transformations. The work shows that errors introduced at Stage 1 progressively shift from raw numerical facts to derived computations, then to narrative prose, and finally to editorially approved conclusions as detectability degrades near-irreversibly. A Markov model over four states quantifies escape probabilities (24.6%, 48.3%, 89.3%), supported by 346 injected hallucinations in a finance analysis pipeline with measurable survival rates.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Introduction

Background and Related Work

Single-Agent Hallucination Detection

Hallucination Snowball Effect

Multi-agent Pipeline Experiments

Markov Model and Transition Probabilities

Optimal Verification Resource Allocation

The Hallucination Snowball: Modeling Error Propagation as State Transitions in Multi-Agent LLM Pipelines Prabhjot Singh 1 2 Bhushan Pawar 2 arXiv :2608 . 14588v 1 [ cs .AI] 22 Jun 2026 Abstract 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 → Derived → Narrative → 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 ∼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 S1 →S2 first, where 75 .4% of hallucinations are still catchable, not at S3 →S4 where 89.3% have already escaped. 1The University of Texas at Austin, Austin, TX, USA 2RediMinds Inc., USA. Correspondence to: Prabhjot Singh <[prabhjot.singh@utexas.edu](prabhjot.singh@utexas.edu) >, Bhushan Pawar <[bhushan.pawar@rediminds.com](bhushan.pawar@rediminds.com) >. Accepted at the FAGEN Workshop (Failure Modes in Agentic AI), ICML 2026 (43rd International Conference on Machine Learning), Seoul, South Korea, 2026 . Copyright 2026 by the author(s) . 1. Introduction Multi-agent LLM pipelines have become the dominant architecture for complex, high-stakes AI tasks. Frameworks like LangGraph (LangChain AI, 2024), AutoGen (Wu et al., 2023), and CrewAI (CrewAI Inc., 2024) make it trivial to chain specialized agents sequentially: Researcher → Analyst → Writer → Reviewer. Each agent receives only its predecessor’s text output with no provenance metadata, no confidence scores, and no access to source documents. Every agent trusts its predecessor completely. This architectural choice has a measurable and surprisingly severe consequence. Consider a hallucination introduced at Stage 1: “COGS was $71.2B”(true value: $63.1B) . By Stage 2, the Analyst computes “an 8.3% YoY increase,” a derived claim built on a fabricated base. By Stage 3, the Writer produces “Boeing faced significant cost headwinds, with COGS rising 8.3% to $71.2B, reflecting supply chain disruptions.” By Stage 4, the Reviewer approves it as “well-structured analysis with sound reasoning,” having been given no source documents and only able to check internal consistency. The hallucination has been laundered completely. Prior work frames hallucination mitigation as finding a better detector (Min et al., 2023 ; Dhuliawala et al., 2024 ; Songet al., 2024 ; Wei et al., 2024) . We show this framing is fundamentally incomplete for sequential multi-agent systems. gpt-4o detection drops from 72.0% at Stage 1 to 50.9% at Stage 4, a 21.1 pp decay regardless of detector sophistication. End-of-pipeline verification achieves only 2.3 pp improvement over no verification (from 60.7% to 58.4% survival) because the information required to verify the original claim is structurally destroyed by upstream transformatio

FAQs

What is the hallucination snowball effect described in the paper?

Hallucinations introduced early in a sequential multi-agent pipeline become increasingly irreversible as they are transformed across stages, moving from raw facts to derived claims, then narrative, and finally undetectable outputs.

How is hallucination propagation modeled?

The paper formalizes propagation as a first-order Markov process over four states, with empirically measured per-boundary escape probabilities.

Where should verification be applied for best results?

Verification should be invested earlier, between Stage 1 and Stage 2, because many hallucinations are still catchable then, while later boundaries exhibit much higher survival.

Sep 4, 2026

Uploaded by Emma Mercer

The Hallucination Snowball - Modeling Error Propagation as State Transitions in Multi-Agent LLM Pipelines_Page 1

Open in WPS to view all 10 pages