Aegis

Automated Error Generation and Attribution for Multi-Agent Systems

Fanqi Kong1,2*, Ruijie Zhang3*, Huaxiao Yin3, Guibin Zhang5, Xiaofei Zhang4,
Ziang Chen4, Zhaowei Zhang1, Xiaoyuan Zhang1, Song-Chun Zhu1,2,4, Xue Feng2†
1PKU 2BIGAI 3UCAS 4THU 5NUS
*Equal contribution †Corresponding author

A novel framework that creates large-scale datasets of realistic MAS failures with verifiable ground-truth labels

Read Paper View Code Dataset

Abstract

As Multi-Agent Systems (MAS) become increasingly autonomous and complex, understanding their error modes is critical for ensuring their reliability and safety. However, research in this area has been severely hampered by the lack of large-scale, diverse datasets with precise, ground-truth error labels. To address this bottleneck, we introduce Aegis, a novel framework for Automated Error Generation and attribution for Multi-Agent Systems.


By systematically injecting controllable and traceable errors into initially successful trajectories, we create a rich dataset of realistic failures. This is achieved using a context-aware, LLM-based adaptive manipulator that performs sophisticated attacks like prompt injection and response corruption to induce specific, predefined error modes. We demonstrate the value of our dataset by exploring three distinct learning paradigms for the error attribution task: Supervised Fine-Tuning, Reinforcement Learning, and Contrastive Learning.

Method Overview

Aegis follows a principled three-stage pipeline to generate high-quality error data and supports multiple learning paradigms for robust error attribution.

Aegis Framework Overview

The Aegis framework automatically generates labeled failures by taking successful multi-agent trajectories and applying controlled, context-aware error injections, enabling three distinct learning methods for error attribution.

Results

Performance Highlights

  • Aegis-SFT achieves 26.51 average score
  • 2× improvement over base models
  • 9,533 trajectories with 24,843 error instances
  • 6 MAS frameworks × 6 task domains

Performance by Domain

Performance Radar Chart

Aegis-SFT (orange) consistently outperforms all baseline models across different task domains and MAS frameworks.

Complete Results on Aegis-Bench

Model Pair Agent Error Avg.
μF1 MF1 μF1 MF1 μF1 MF1
Random Baseline 0.33 0.21 4.54 3.56 11.23 11.15 4.08
Small-Scale Models
DCL (Ours) 8.33 5.30 22.93 20.23 24.73 27.70 12.61
Medium-Scale Models
Qwen2.5-7B-Instruct 5.02 2.52 27.55 14.49 14.96 11.36 12.43
+ SFT 5.05 2.80 60.03 22.70 19.61 16.90 17.99
+ GRPO 7.11 2.77 35.43 14.86 17.21 10.54 14.87
Qwen2.5-14B-Instruct 5.47 2.20 35.78 12.71 20.24 5.91 13.99
+ SFT (Aegis-SFT) 16.62 9.99 76.53 47.97 27.53 27.66 26.51
+ GRPO (Aegis-GRPO) 6.84 2.55 49.74 18.38 21.19 16.10 18.41
Qwen3-8B-Non-Thinking 3.96 1.40 21.34 8.16 15.81 13.89 10.12
+ SFT 9.68 5.73 64.79 38.96 20.37 20.36 21.41
+ GRPO 6.94 2.82 45.91 17.39 20.89 15.15 17.15
Qwen3-8B-Thinking 4.42 1.52 34.63 9.01 17.48 14.31 13.06
+ GRPO 4.41 1.66 36.11 15.73 17.94 12.03 17.58
Large-Scale Models
Qwen2.5-72B-Instruct 5.60 2.20 37.46 14.51 17.72 16.58 15.01
gpt-oss-120b 6.53 1.71 38.58 5.53 20.38 12.05 17.07
GPT-4.1 7.44 2.27 37.48 11.12 20.65 15.75 15.27
GPT-4o-mini 5.76 1.63 38.54 14.72 19.95 16.02 15.83
o3 7.86 2.27 40.31 23.27 22.37 16.76 20.24
Gemini-2.5-Flash 6.99 2.76 42.02 16.45 23.47 19.85 19.55
Gemini-2.5-Pro 6.96 2.88 41.32 16.15 19.93 16.29 18.35
Claude-Sonnet-4 7.68 2.34 40.73 15.51 21.21 16.55 18.16

Resources

Paper

Read the full paper on arXiv with detailed methodology and comprehensive experiments

Code

Access the complete codebase, including data generation pipeline and training scripts

Dataset

Download the Aegis dataset with 9,533 annotated error trajectories

Models

Pre-trained models on Hugging Face for error attribution in multi-agent systems

Benchmark

Aegis-Bench evaluation suite for systematic MAS error attribution

Documentation

Comprehensive guide for using Aegis in your own multi-agent systems

Citation

If you find Aegis useful for your research, please cite our paper:

@article{kong2025aegis, title={Aegis: Automated Error Generation and Attribution for Multi-Agent Systems}, author={Kong, Fanqi and Zhang, Ruijie and Yin, Huaxiao and Zhang, Guibin and Zhang, Xiaofei and Chen, Ziang and Zhang, Zhaowei and Zhang, Xiaoyuan and Zhu, Song-Chun and Feng, Xue}, journal={arXiv preprint arXiv:2509.14295}, year={2025} }

Click here to copy citation to clipboard