On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents

TL;DR

This study analyzes how different MAS structures (Linear, Flat, Hierarchical) resist faults introduced to agents via AutoTransform and AutoInject across four tasks. It demonstrates that hierarchical structures yield the best resilience, with minimal performance drops, and shows that fault frequency (number of faulty messages) has a larger impact than fault severity within messages. The authors further introduce Challenger and Inspector defenses that can recover up to 96.4% of lost performance, providing a practical approach to improving robustness. Overall, the work highlights the importance of organizational structure in MAS resilience and offers concrete methods to simulate faults and defend against them, with code and data publicly available.

Abstract

Large language model-based multi-agent systems have shown great abilities across various tasks due to the collaboration of expert agents, each focusing on a specific domain. However, the impact of clumsy or even malicious agents--those who frequently make errors in their tasks--on the overall performance of the system remains underexplored. This paper investigates: (1) What is the resilience of various system structures (e.g., A$\rightarrow$B$\rightarrow$C, A$\leftrightarrow$B$\leftrightarrow$C) under faulty agents, on different downstream tasks? (2) How can we increase system resilience to defend against these agents? To simulate faulty agents, we propose two approaches--AutoTransform and AutoInject--which introduce mistakes into the agents' responses. Experiments on four downstream tasks using six systems show that the "hierarchical" structure, i.e., A$\rightarrow$(B$\leftrightarrow$C), exhibits superior resilience with the lowest performance drop of 5.5%, compared to 10.5% and 23.7% of other two structures. To further improve resilience, we introduce (1) Challenger, that introduces a mechanism for each agent to challenge others' outputs, and (2) Inspector, an additional agent to review and correct messages, recovering up to 96.4% errors made by faulty agents. Our code and data are available at https://github.com/CUHK-ARISE/MAS-Resilience.

Figures (6)

• Figure 1: We focus on the overall impact of faulty agents on the performance of diverse system structures across various tasks.
• Figure 2: Overview of our error-introducing methods. (a) Task information. (b) Multi-agent collaboration system without faulty agents. (c) AutoTransform modifies agent's profile to turn it into faulty while preserving original functionalities. (d) AutoInject intercepts messages between agents and adds errors into the messages.
• Figure 3: The performance of various system structures with the two error-introducing methods, with results averaged across all four tasks.
• Figure 4: The performance of various tasks with the two error-introducing methods, with results averaged across three system structures (all six multi-agent systems).
• Figure 5: The performance of all six GPT-3.5-based multi-agent systems in code generation, using AutoInject to introduce errors.