Elamur Research⌕
☰
home / papers / 2026 / MADRA: Multi-Agent Debate for Risk-Aware Embodied Planning

MADRA: Multi-Agent Debate for Risk-Aware Embodied Planning

Paper · 2026 · JunJian Wang, Lidan Zhao and Xi Sheryl Zhang

Summary

LLM-driven household robots can produce plans that read fluently but are physically dangerous, such as heating a book in a microwave. The authors call this "Embodied Semantic Hallucination." Common fixes like RLHF or safety prompting tend to make agents refuse harmless requests, which trades usefulness for safety.

MADRA is a training-free, plug-in layer for any LLM backbone, framed as System-2 deliberation. Several debate agents each judge whether an instruction is safe and give risk categories and reasons. A separate Critical Agent then grades each argument using criteria derived from the Toulmin argumentation model. These scores steer the next debate round, which counters agents' tendency to conform to the majority. The debate stops when the agents agree. If no consensus is reached within the round limit, majority voting decides.

The authors also release SafeAware-VH, a VirtualHome benchmark for this task. Tested on it and on SafeAgentBench (AI2-THOR), MADRA rejects over 90% of unsafe tasks while keeping refusals of safe tasks low. It outperforms Chain-of-Thought and single-agent reflection baselines.

Key points

  • SafeAware-VH benchmark: 800 expert-annotated instructions labeled with a taxonomy of physical risks. It includes an Adversarial Safe Subset designed to measure over-rejection (ORR) alongside unsafe rejection (URR).
  • Critical Agent scoring: arguments are weighted by Logical Soundness (30%), Risk Identification (30%), Evidence Quality (30%) and Clarity (10%). The scoring penalizes slippery-slope reasoning and invented environment details.
  • Backbones tested: GPT-4o, GPT-3.5, Qwen, DeepSeek and Llama3. Qwen and DeepSeek perform close to GPT-4o, suggesting cheaper models are viable.
  • Ablations: 95% of debates converge within 3 rounds. Removing the Critical Agent makes results unstable and encourages herd behavior. Including it improves performance by about 5%.
  • Limitations: evaluation uses text-based simulation only, results depend on the LLM's own hazard knowledge, and deliberation adds latency.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🏛️Chinese Academy of SciencesBeijing · China194 works · 23,906 citations🔬Institute of AutomationBeijing · China70 works · 14,251 citations🎓University of Chinese Academy of SciencesBeijing · China147 works · 8,480 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksEngineering · 133 worksSafety, Risk, Reliability and Quality · 10 works

Cite

Cite as JunJian Wang, Lidan Zhao and Xi Sheryl Zhang. 2026. MADRA: Multi-Agent Debate for Risk-Aware Embodied Planning. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.340

@inproceedings{wang2026madra,
  title = {{MADRA: Multi-Agent Debate for Risk-Aware Embodied Planning}},
  author = {JunJian Wang, Lidan Zhao and Xi Sheryl Zhang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.340},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.340}
}

References

28 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Grady Booch, Francesco Fabiano, Lior Horesh, Kiran Kate, Jonathan Lenchner, Nick Linck, Andreas Loreggia, Keerthiram Murgesan, Nicholas Mattei, Francesca Rossi, and 1 others. 2021. Thinking fast and slow in ai. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 15042–… find
  2. Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, and 1 others. 2023. Do as i can, not as i say: Grounding language in robotic affordances. In Conference on robot learning, pages 287–318. PMLR.
  3. Jae-Woo Choi, Youngwoo Yoon, Hyobin Ong, Jaehong Kim, and Minsu Jang. 2024. Lota-bench: Benchmarking language-oriented task planners for embodied agents. arXiv preprint arXiv:2402.08178. — [arXiv:2402.08178](https://arxiv.org/abs/2402.08178) find
  4. Inner monologue: Embodied reasoning through planning with language models. In Conference on Robot Learning, pages 1769–1782. PMLR. find
  5. Yuting Huang, Leilei Ding, Zhipeng Tang, Tianfu Wang, Xinrui Lin, Wuyang Zhang, Mingxiao Ma, and Yanyong Zhang. 2025. A framework for benchmarking and aligning task-planning safety in llm-based embodied agents. arXiv preprint arXiv:2504.14650. — [arXiv:2504.14650](https://arxiv.org/abs/2504.14650) find
  6. Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, and 1 others. 2023. Llama guard: Llm-based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674.
  7. Daniel Kahneman. 2011. Thinking, fast and slow. Farrar, Straus and Giroux. find
  8. Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023. Camel: Communicative agents for" mind" exploration of large language model society. Advances in Neural Information Processing Systems, 36:51991–52008.
  9. Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. 2023. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493–9500. IEEE. find
  10. Aishan Liu, Zonghao Ying, Le Wang, Junjie Mu, Jinyang Guo, Jiakai Wang, Yuqing Ma, Siyuan Liang, Mingchuan Zhang, Xianglong Liu, and 1 others. 2025. Agentsafe: Benchmarking the safety of embodied agents on hazardous instructions. arXiv preprint arXiv:2506.14697. — [arXiv:2506.14697](https://arxiv.org/abs/2506.14697) find
  11. Xiaoya Lu, Zeren Chen, Xuhao Hu, Yijin Zhou, Weichen Zhang, Dongrui Liu, Lu Sheng, and Jing Shao. 2025. Is-bench: Evaluating interactive safety of vlm-driven embodied agents in daily household tasks. arXiv preprint arXiv:2506.16402. — [arXiv:2506.16402](https://arxiv.org/abs/2506.16402) find
  12. Marvin Minsky. 1986. Society of mind. Simon and Schuster. find
  13. Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. 2023. Improving factuality and reasoning in language models through multiagent debate. In Forty-first International Conference on Machine Learning.
  14. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36:8634–8652.
  15. Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. 2023. Critic: Large language models can self-correct with tool-interactive critiquing. arXiv preprint arXiv:2305.11738.
  16. Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. 2023. Progprompt: program generation for situated robot task planning using large language models. Autonomous Robots, 47(8):999–1012. find
  17. Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, and 1 others. 2023. find
  18. Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. 2023. 6860 9 find
  19. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2998–3009.
  20. Stephen E Toulmin. 2003. The uses of argument. Cambridge university press. find
  21. Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, and 1 others. 2024. A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345.
  22. Zhiyuan Weng, Guikun Chen, and Wenguan Wang. 2025. Do as we do, not as you think: the conformity of large language models. arXiv preprint arXiv:2501.13381. — [arXiv:2501.13381](https://arxiv.org/abs/2501.13381) find
  23. Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, and 1 others. 2025. The rise and potential of large language model based agents: A survey. Science China Information Sciences, 68(2):121101.
  24. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR).
  25. Sheng Yin, Xianghe Pang, Yuanzhuo Ding, Menglan Chen, Yutong Bi, Yichen Xiong, Wenhao Huang, Zhen Xiang, Jing Shao, and Siheng Chen. 2024. Safeagentbench: A benchmark for safe task planning of embodied llm agents. arXiv preprint arXiv:2412.13178. — [arXiv:2412.13178](https://arxiv.org/abs/2412.13178) find
  26. Tongxin Yuan, Zhiwei He, Lingzhong Dong, Yiming Wang, Ruijie Zhao, Tian Xia, Lizhen Xu, Binglin Zhou, Fangqi Li, Zhuosheng Zhang, and 1 others. 2024. R-judge: Benchmarking safety risk awareness for llm agents. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 1467–1490.
  27. Hangtao Zhang, Chenyu Zhu, Xianlong Wang, Ziqi Zhou, Shengshan Hu, and Leo Yu Zhang. 2024. Badrobot: Jailbreaking llm-based embodied ai in the physical world. arXiv preprint arXiv:2407.20242. — [arXiv:2407.20242](https://arxiv.org/abs/2407.20242) find
  28. Mingchen Zhuge, Haozhe Liu, Francesco Faccio, Dylan R Ashley, Robert Csordás, Anand Gopalakrishnan, Abdullah Hamdi, Hasan Abed Al Kader Hammoud, Vincent Herrmann, Kazuki Irie, and 1 others. 2025. Mindstorms in natural language-based societies of mind. Computational Visual Media, 11(1):29–81. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-madra-multi-agent-debate-risk/