Elamur Research
☰
home / papers / 2023 / ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate

ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "arXiv non-exclusive distribution license" does not allow redistributing the text itself. Original — arxiv.org/pdf/2308.07201

Summary

Judging the quality of generated text is costly when done by people. N-gram metrics such as BLEU, Rouge and METEOR are cheap but align poorly with human opinion on open-ended outputs. Using a single LLM as a judge narrows this gap but does not yet reach human-level quality. ChatEval responds by turning evaluation into a group discussion. Several LLM agents form a referee panel and debate candidate responses before giving a verdict, mirroring the common practice of having multiple annotators work together. Each agent gets its own persona, and the authors test three schemes for sharing discussion history. Rather than forcing consensus, final labels are obtained by majority vote or averaged scores. The panels were built from GPT-4 or ChatGPT and tested on FairEval (open-ended QA) and Topical-Chat (dialogue response generation). In both cases, the debating panels matched human judgments more closely than single-agent evaluators, in both accuracy and correlation. Assigning every agent the same role lowered performance, so persona diversity is essential. The released framework offers a scalable, more human-like route to automatic evaluation.

Key points

  • Benchmarks: FairEval uses 80 open-ended questions from Chiang et al. (2023), comparing Vicuna-13B and ChatGPT answers. Its human labels come from Wu et al. (2023), with three annotators and a majority vote. Topical-Chat covers dialogue response generation.
  • Agents are homogeneous groups of either GPT-4 or ChatGPT (GPT-3.5-turbo), run at temperature 0. Mixed-model groups are left for future work.
  • The three communication strategies are:
    • One-By-One: agents speak in a fixed order.
    • Simultaneous-Talk: agents respond in parallel, which removes order effects.
    • Simultaneous-Talk-with-Summarizer: an extra LLM condenses each round into the shared history.
  • Role prompts are adapted from Wu et al. (2023): General Public, Critic, News Author, Psychologist and Scientist. Using identical role descriptions for all agents degrades results.
  • Code is available at github.com/chanchimin/ChatEval and is built on the AgentVerse library.

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

Fields: AI & machine learning

Cited by

297 papers of our library cite it (of the 140673 whose bibliographies we read); 11 of them have a page here, 286 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2308.07201 · arXiv 2308.07201

@misc{chan2023chateval,
  title = {{ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate}},
  author = {Chan, Chi-Min and Chen, Weize and Su, Yusheng and Yu, Jianxuan and Xue, Wei and Zhang, Shanghang and Fu, Jie and Liu, Zhiyuan},
  year = {2023},
  doi = {10.48550/arxiv.2308.07201},
  eprint = {2308.07201},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2308.07201}
}

References

46 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yushi Bai, Jiahao Ying, Yixin Cao, Xin Lv, Yuze He, Xiaozhi Wang, Jifan Yu, Kaisheng Zeng, Yijia Xiao, Haozhe Lyu, et al. Benchmarking foundation models with language-model-as-an-examiner. arXiv preprint arXiv:2306.04181, 2023. find
  2. Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pp. 65–72, 2005. find our card
  3. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020. find
  4. Chris Callison-Burch. Fast, cheap, and creative: Evaluating translation quality using amazon’s mechanical turk. In Proceedings of the 2009 conference on empirical methods in natural language processing, pp. 286–295, 2009. find
  5. Asli Celikyilmaz, Elizabeth Clark, and Jianfeng Gao. Evaluation of text generation: A survey. CoRR, abs/2006.14799, 2020. URL https://arxiv.org/abs/2006.14799. ↗
  6. Cheng-Han Chiang and Hung-yi Lee. Can large language models be an alternative to human evaluations? arXiv preprint arXiv:2305.01937, 2023. find
  7. Acc. (%) Kap. 0.32 0.36 0.30 0.34 Acc. (%) 0.28 Kap. Kap. 0.32 0.26 0.30 0.24 0.28 3 Discussion Turns (b) Acc. and Kap. vs Discussion Turns Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-… find
  8. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. find
  9. Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325, 2023. find our card
  10. Jinlan Fu, See-Kiong Ng, Zhengbao Jiang, and Pengfei Liu. Gptscore: Evaluate as you desire. arXiv preprint arXiv:2302.04166, 2023. find
  11. Mingqi Gao, Jie Ruan, Renliang Sun, Xunjian Yin, Shiping Yang, and Xiaojun Wan. Human-like summarization evaluation with chatgpt. arXiv preprint arXiv:2304.02554, 2023. find
  12. Karthik Gopalakrishnan, Behnam Hedayatnia, Qinlang Chen, Anna Gottardi, Sanjeev Kwatra, Anushree Venkatesh, Raefer Gabriel, and Dilek Hakkani-Tür. Topical-chat: Towards knowledge-grounded open-domain conversations. 2019. find
  13. Marzena Karpinska, Nader Akoury, and Mohit Iyyer. The perils of using mechanical turk to evaluate open-ended text generation. arXiv preprint arXiv:2109.06835, 2021. find
  14. Grzegorz Kondrak. N-gram similarity and distance. In International symposium on string processing and information retrieval, pp. 115–126. Springer, 2005. find
  15. Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for” mind” exploration of large scale language model society. arXiv preprint arXiv:2303.17760, 2023a. find our card
  16. Ruosen Li, Teerth Patel, and Xinya Du. Prd: Peer rank and discussion improve large language model based evaluations. arXiv preprint arXiv:2307.02762, 2023b. find
  17. Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Zhaopeng Tu, and Shuming Shi. Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118, 2023. find our card
  18. Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pp. 74–81, 2004. find
  19. Ruibo Liu, Ruixin Yang, Chenyan Jia, Ge Zhang, Denny Zhou, Andrew M Dai, Diyi Yang, and Soroush Vosoughi. Training socially aligned language models in simulated human society. arXiv preprint arXiv:2305.16960, 2023a. find
  20. Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. Gpteval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634, 2023b. find
  21. Andrea I Luppi, Pedro AM Mediano, Fernando E Rosas, Negin Holland, Tim D Fryer, John T O’Brien, James B Rowe, David K Menon, Daniel Bor, and Emmanuel A Stamatakis. A synergistic core for human brain evolution and cognition. Nature Neuroscience, 25(6):771–782, 2022. find
  22. Zhao Mandi, Shreeya Jain, and Shuran Song. Roco: Dialectic multi-robot collaboration with large language models. arXiv preprint arXiv:2307.04738, 2023. — [arXiv:2307.04738](https://arxiv.org/abs/2307.04738) find
  23. Mary L McHugh. Interrater reliability: the kappa statistic. Biochemia medica, 22(3):276–282, 2012. find
  24. Shikib Mehri and Maxine Eskenazi. Usr: An unsupervised and reference free evaluation metric for dialog generation. arXiv preprint arXiv:2005.00456, 2020. find our card
  25. Jekaterina Novikova, Ondřej Dušek, Amanda Cercas Curry, and Verena Rieser. Why we need new evaluation metrics for NLG. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 2241–2252, Copenhagen, Denmark, September 2017. Association for Computational Lingui… find
  26. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730–27744, 2022. find
  27. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pp. 311–318, 2002. find our card
  28. Joon Sung Park, Joseph C O’Brien, Carrie J Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442, 2023. find
  29. Chen Qian, Xin Cong, Cheng Yang, Weize Chen, Yusheng Su, Juyuan Xu, Zhiyuan Liu, and Maosong Sun. Communicative agents for software development. arXiv preprint arXiv:2307.07924, 2023. find our card
  30. Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover’s distance as a metric for image retrieval. International journal of computer vision, 40:99–121, 2000. find
  31. Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207, 2021. find our card
  32. Thibault Sellam, Dipanjan Das, and Ankur P Parikh. Bleurt: Learning robust metrics for text generation. arXiv preprint arXiv:2004.04696, 2020. find
  33. Chenhui Shen, Liying Cheng, Yang You, and Lidong Bing. Are large language models good evaluators for abstractive summarization? arXiv preprint arXiv:2305.13091, 2023. find
  34. Amanda Stent, Matthew Marge, and Mohit Singhai. Evaluating evaluation methods for generation in the presence of variation. In International conference on intelligent text processing and computational linguistics, pp. 341–351. Springer, 2005. find
  35. Chris Van Der Lee, Albert Gatt, Emiel Van Miltenburg, Sander Wubben, and Emiel Krahmer. Best practices for the human evaluation of automatically generated text. In Proceedings of the 12th International Conference on Natural Language Generation, pp. 355–368, 2019. find
  36. Jiaan Wang, Yunlong Liang, Fandong Meng, Haoxiang Shi, Zhixu Li, Jinan Xu, Jianfeng Qu, and Jie Zhou. Is chatgpt a good nlg evaluator? a preliminary study. arXiv preprint arXiv:2303.04048, 2023a. find our card
  37. Peiyi Wang, Lei Li, Liang Chen, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926, 2023b. find
  38. Yidong Wang, Zhuohao Yu, Zhengran Zeng, Linyi Yang, Cunxiang Wang, Hao Chen, Chaoya Jiang, Rui Xie, Jindong Wang, Xing Xie, et al. Pandalm: An automatic evaluation benchmark for llm instruction tuning optimization. arXiv preprint arXiv:2306.05087, 2023c. find our card
  39. Zhenhailong Wang, Shaoguang Mao, Wenshan Wu, Tao Ge, Furu Wei, and Heng Ji. Unleashing cognitive synergy in large language models: A task-solving agent through multi-persona self-collaboration. arXiv preprint arXiv:2307.05300, 2023d. find our card
  40. Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. find
  41. Anita Williams Woolley, Christopher F Chabris, Alex Pentland, Nada Hashmi, and Thomas W Malone. Evidence for a collective intelligence factor in the performance of human groups. science, 330(6004):686–688, 2010. find
  42. Ning Wu, Ming Gong, Linjun Shou, Shining Liang, and Daxin Jiang. Large language models are diverse role-players for summarization evaluation. arXiv preprint arXiv:2303.15078, 2023. find
  43. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. find
  44. Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M Meyer, and Steffen Eger. Moverscore: Text generation evaluating with contextualized embeddings and earth mover distance. arXiv preprint arXiv:1909.02622, 2019. find our card
  45. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023. find
  46. Ming Zhong, Yang Liu, Da Yin, Yuning Mao, Yizhu Jiao, Pengfei Liu, Chenguang Zhu, Heng Ji, and Jiawei Han. Towards a unified multi-dimensional evaluator for text generation. arXiv preprint arXiv:2210.07197, 2022. find