Task Decomposition Blindness in Multi-Agent LLM Orchestration: Cross-Domain Safety Failures

Read full text Summarise with VIRAT
Authors: HIROKI FUKUI
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This study examines whether multi-agent LLM orchestration systems systematically miss cross-section inconsistencies in specification documents that a single LLM instance can detect. We test this across four high-stakes domains (healthcare, finance, law, software) using five agent configurations that decompose the "cost of decomposition" into four incremental steps: information splitting, multi-agent coordination, hierarchy, and orchestrator invisibility. Pilot data (N=20, excluded from confirmatory analysis) showed Solo ETR=100% vs Multi-agent ETR=37-50%, confirming the phenomenon. This study is part of the SociA project (Alignment Psychopathology); Series O (osf.io/tw6mz) established that invisible orchestration increases internal dissociation, and Series O-2 extends this by measuring downstream fatal output generation.

Fields of Study & Subjects

Ethics and Social Impacts of AI Multi-Agent Systems and Negotiation Natural Language Processing Techniques

Publication details

Publication date
1 January 2026
Language
EN
Document type
Preprint