Research map / Multi-agent systems and games / LLM agents
On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies how faulty or clumsy agents affect six LLM multi-agent systems (linear, flat, hierarchical) across four tasks, introduces AutoTransform and AutoInject for simulating errors, and proposes Challenger and Inspector defenses.
Problem
LLM-based multi-agent systems are vulnerable to clumsy or malicious agents that make stealthy errors, yet how system structure affects resilience to such faulty agents on general downstream tasks is underexplored.
Approach
Two error-introduction methods—AutoTransform (rewrites an agent's profile to make it faulty while preserving function) and AutoInject (directly injects controlled errors into inter-agent messages)—are used to compare structural resilience, and two defenses (a Challenger profile addition and an Inspector agent that reviews messages) are proposed to recover performance.
Key results
- Hierarchical structure shows the smallest average performance drop (5.51) versus flat (10.54) and linear (23.72) under faulty agents.
- Code generation is most sensitive, with an average drop of 22.56 points, while translation is least affected (4.70).
- Challenger combined with Inspector recovers 96.4% of performance lost by faulty agents on Self-collab (code generation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a hierarchical rather than linear collaboration structure | GPT-3.5 and GPT-4o multi-agent systems · All four tasks | average performance drop: −23.72 to −5.51 | Helped |
| Introduce faulty agents using AutoTransform | GPT-3.5 multi-agent systems · All four tasks | average performance: 54.33 -> 42.91 | Hurt |
| Inject errors into messages using AutoInject | GPT-3.5 multi-agent systems · All four tasks | average performance: 54.33 -> 42.21 | Hurt |
| Introduce faulty agents using AutoTransform | GPT-4o multi-agent systems · All four tasks | average performance: 67.50 -> 49.28 | Hurt |
| Inject errors into messages using AutoInject | GPT-4o multi-agent systems · All four tasks | average performance: 67.50 -> 56.22 | Hurt |
| Collaborate using multiple agents rather than one | GPT-3.5 · HumanEval, CIAR, CommonMT, FairEval | average performance: 48.02 -> 52.78 | Helped |
| Collaborate using multiple agents rather than one | GPT-4o · HumanEval, CIAR, CommonMT, FairEval | average performance: 60.49 -> 65.78 | Helped |
| Inject semantic errors rather than syntactic errors | Six multi-agent systems · HumanEval | accuracy: 64.73 -> 39.15; 64.73 -> 47.78 | Mixed |
| Increase the proportion of messages containing errors (ablation) | AutoInject systems · HumanEval | average performance: 60.57 -> 52.94 -> 47.53 | Hurt |
| Increase errors within each message from $P_e=0.4$ to $P_e=0.6$ when $P_m=0.2$ (ablation) | AutoInject systems · HumanEval | average performance: 53.05 -> 53.75 | Mixed |
| Add Challenger (ablation) | Self-collab with AutoTransform errors · HumanEval | accuracy: 43.3 -> 70.7 | Helped |
| Add Inspector (ablation) | Self-collab with AutoTransform errors · HumanEval | accuracy: 43.3 -> 74.4 | Helped |
Surprising findings
- Injecting errors sometimes improves performance, e.g., MAD on text evaluation with GPT-3.5 gains up to 12.1%, attributed to double-checking and divergent thinking.
- Current LLMs prioritize natural-language comments over code; a distraction comment can make systems accept buggy code.
Limitations
- Only one faulty agent is introduced at a time in main experiments; multi-fault settings are explored only briefly in the appendix.
- AutoTransform does not reliably control error rates, so factor analysis relies on AutoInject.
Experimental setup
Datasets: HumanEval CIAR CommonMT FairEval
Benchmarks: HumanEval CIAR CommonMT FairEval
Metrics: Pass@1 accuracy accuracy BLEURT-20 score performance drop
Baselines: MetaGPT Self-collaboration Camel SPP MAD AgentVerse single agent
Builds on: MetaGPT Self-collaboration Camel SPP MAD AgentVerse
Closest papers by meaning
- ResMAS: Resilience Optimization in LLM-based Multi-agent Systems (AAAI 2026 · similarity 0.91)
- Why Do Multi-Agent LLM Systems Fail? (NeurIPS 2025 · similarity 0.88)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.88)
- AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators (NeurIPS 2026 · similarity 0.88)
- AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems? (ICLR 2026 · similarity 0.88)
- DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems (ICLR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.