Multi-agent AI pipelines suffer from 'hallucination snowball' effect
Errors injected at the start of multi-agent LLM pipelines multiply and evade detection as they're transformed from facts into narratives.
A new paper shows that hallucinations in multi-agent LLM pipelines don't just persist—they snowball. Errors injected at Stage 1 transform through four states, becoming nearly undetectable by Stage 4. Detection with gpt-4o 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. Boundary verification at early stages cuts hallucination survival from 58.4% to 16.2% versus end-of-pipeline checking—proving when you verify matters more than whether you verify.
- Hallucinations injected at the start of multi-agent LLM pipelines multiply and become nearly undetectable by the final stage, with detection dropping from 72% to 24%.
- Boundary verification (early-stage gates) cuts hallucination survival from 58.4% to 16.2%, outperforming end-of-pipeline checks by 35x (2.3 pp vs. 16.2 pp improvement).
- The 'hallucination snowball' effect follows a 4-state Markov process (Raw Fact → Derived → Narrative → Invisible) with measured escape probabilities of 24.6%, 48.3%, and 89.3%.
Why It Matters
Early-stage verification in multi-agent pipelines could reduce critical errors by 72% in high-stakes domains like finance or healthcare.