Capture Notes

OpenReview page for ICML 2025 poster.

Why Collected

Relevant to propagation because it stress-tests how faulty or malicious agents degrade multi-agent collaboration under different graph structures and defenses.

Key Metadata

• Published: 2025-05-01
• Last modified: 2025-07-23
• Venue: ICML 2025 poster
• Code: https://github.com/CUHK-ARISE/MAS-Resilience

Collection Summary

The work introduces AutoTransform and AutoInject to inject mistakes into agent responses or roles. It evaluates structures such as chains, flat peer groups, and hierarchies, finding hierarchical structures more robust. It introduces Challenger and Inspector mechanisms to challenge or correct messages.

Security Relevance

• Provides graph topology and faulty-agent propagation evidence.
• Useful as a defense baseline: hierarchical control, Challenger, Inspector, independent review.
• Does not focus on self-evolving memory/skill artifacts, so use as propagation/resilience analog evidence.

Suggested Ingest Focus

• Create evidence around topology-sensitive failure propagation and inspector/challenger mitigation.
• Link to [[04_Research_Questions/RQ - MAS Misevolution Propagation Control]] as indirect support.