The Hallucination Snowball: Modeling Error Propagation as State Transitions in Multi-Agent LLM Pipelines
4.20T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2608.14588.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryPaper formalizes hallucination propagation in sequential multi-agent LLM pipelines as a four-state Markov process (Raw Fact → Derived → Narrative → Invisible). Across 346 injected hallucinations in a 4-agent FinanceBench pipeline, detection drops from 72.0% at Stage 1 to 50.9% at Stage 4, with 23.7% surviving undetected. Boundary gates cut survival to 16.2% versus 58.4% for end-checking alone.
Why it mattersQuantifies a structural flaw in multi-agent chains and prescribes where to place verification gates, with effect sizes generalizable to any n-agent linear pipeline.
Cited by
No citations on record.