Capture Notes

arXiv:2603.04474. Last revised 2026-05-11.

Why Collected

Directly relevant to [[04_Research_Questions/RQ - MAS Misevolution Propagation Control]] because it models how small local errors propagate and amplify through LLM-based multi-agent collaboration.

Key Metadata

• Submitted: 2026-03-04
• Revised: 2026-05-11
• Subject: Multiagent Systems
• DOI: https://doi.org/10.48550/arXiv.2603.04474

Collection Summary

The paper argues that collaborative LLM-MAS can turn minor inaccuracies into system-level false consensus through iterative message dependencies. It proposes a propagation dynamics model that abstracts collaboration as a directed dependency graph, identifies cascade amplification, topological sensitivity, and consensus inertia, and introduces a genealogy-graph governance layer as a message-layer plugin.

Security Relevance

• Strong direct evidence for graph-level propagation in MAS.
• Useful analog for misevolution propagation even though the paper focuses on error cascade rather than self-evolving skill/memory updates.
• Supports metrics such as propagation depth, final infection, graph dependency risk, and genealogy-based containment.

Suggested Ingest Focus

• Create evidence around error cascade as a MAS propagation primitive.
• Link to [[04_Research_Questions/RQ - MAS Misevolution Propagation Control]] and [[04_Research_Questions/RQ - Agentic Web Protocol Trust Boundaries]].
• Distinguish source claim from wiki inference: paper studies error propagation, not necessarily persistent evolution artifact poisoning.