Elamur Research⌕
☰
home / papers / 2023 / Examining Inter-Consistency of Large Language Models Collaboration: An In-depth Analysis via Debate

Examining Inter-Consistency of Large Language Models Collaboration: An In-depth Analysis via Debate

Paper · 2023 · Kai Xiong, Xiao Ding, Yixin Cao et al.

Summary

Work on LLM consistency usually examines a single model. This paper asks whether several LLMs can work together and agree on a shared answer. The authors use multi-choice commonsense reasoning as a testbed because disagreement between models can be measured exactly. They define a metric, INCON, which is the share of samples where at least two models give different answers. They also propose FORD, a formal debate framework with three steps. First, each model picks an answer and justifies it. Second, on samples where the models disagree, they take turns rebutting or conceding. Third, a judge summarizes the exchange and decides any unresolved cases. Three settings are tested: debates between models of similar strength, debates between unequal models, and roundtable debates with more than two participants. Across seven datasets and six LLMs, models often disagree, but debate usually leads them to agreement and can improve accuracy. When models are unequal, the stronger one tends to prevail, although a stubborn weaker model can drag results down. Using GPT-4 as judge improves outcomes. These findings inform the design of multi-agent LLM systems.

Key points

  • Datasets: seven multi-choice commonsense benchmarks were used:
    • αNLI (1,507 items)
    • CSQA (1,221)
    • COPA (500)
    • e-CARE (2,122)
    • Social IQa (1,935)
    • PIQA (1,838)
    • StrategyQA (2,290)
  • Models: six LLMs were tested: ChatGPT (gpt-3.5-turbo), ChatGPT-0301, GPT-4, Davinci-003 (text-davinci-003), LLaMA-13B, and Vicuna-13B.
  • Baselines: FORD is compared with single models and two collaboration baselines:
    • Col-S picks one model at random when they disagree.
    • Col-H counts an answer as correct only if all models are right.
  • Example result (ChatGPT vs. Davinci-003 on αNLI): FORD scores 81.35. This beats ChatGPT alone (77.17), Davinci-003 alone (79.96), Col-S (78.75), and Col-H (69.14).
  • Other findings:
    • Chat and text-completion models disagree substantially, even when they share the same base model.
    • Newer GPT-3.5 versions do not fully replace the abilities of older ones.
    • Code and data are released at github.com/Waste-Wood/FORD.

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

🎓Harbin Institute of TechnologyHarbin · China76 works · 27,918 citations🎓Singapore Management UniversitySingapore · Singapore9 works · 2,130 citations

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 9,086 worksArtificial Intelligence · 8,889 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 worksComputer Vision and Pattern Recognition · 1,282 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 1,343 works

Cite

Cite as Kai Xiong, Xiao Ding, Yixin Cao, Ting Liu and Bing Qin. 2023. Examining Inter-Consistency of Large Language Models Collaboration: An In-depth Analysis via Debate. Findings of the Association for Computational Linguistics: EMNLP 2023.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2023.findings-emnlp.508

@inproceedings{xiong2023examining,
  title = {{Examining Inter-Consistency of Large Language Models Collaboration: An In-depth Analysis via Debate}},
  author = {Kai Xiong, Xiao Ding, Yixin Cao, Ting Liu and Bing Qin},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2023},
  year = {2023},
  doi = {10.18653/v1/2023.findings-emnlp.508},
  url = {https://doi.org/10.18653/v1/2023.findings-emnlp.508}
}

References

36 references, 18 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Philip LaVerne Bell. 1998. Designing for students’ science learning using argumentation and classroom debate. University of California, Berkeley. find
  2. Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman, Hannah Rashkin, Doug Downey, Wen-tau Yih, and Yejin Choi. 2019. Abductive commonsense reasoning. In International Conference on Learning Representations.
  3. Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. 2020. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 7432–7439.
  4. Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. 2023. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712.
  5. Bibo Cai, Xiao Ding, Zhouhao Sun, Bing Qin, Ting Liu, Lifeng Shang, et al. 2023. Self-supervised logic induction for explainable fuzzy temporal commonsense reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 12580– 12588. find
  6. Mark R Chassin, Robert W Galvin, et al. 1998. The urgent need to improve health care quality: Institute of medicine national roundtable on health care quality. Jama, 280(11):1000–1005. find
  7. Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality. find
  8. Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416.
  9. Bhavana Dalvi, Peter Jansen, Oyvind Tafjord, Zhengnan Xie, Hannah Smith, Leighanna Pipatanangkura, and Peter Clark. 2021. Explaining answers with entailment trees. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7358–7370.
  10. Li Du, Xiao Ding, Kai Xiong, Ting Liu, and Bing Qin. 2022a. e-care: a new dataset for exploring explainable causal reasoning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 432–446. find
  11. Li Du, Xiao Ding, Kai Xiong, Ting Liu, and Bing Qin. 2022b. Enhancing pretrained language models with structured commonsense knowledge for textual inference. Knowledge-Based Systems, 254:109488. find
  12. Yao Fu, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. 2023. Complexity-based prompting for multi-step reasoning. International Conference on Learning Representations.
  13. Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346– 361. find
  14. Andrew Gordon, Zornitsa Kozareva, and Melissa Roemmele. 2012. SemEval-2012 task 7: Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In *SEM 2012: The First Joint Conference on Lexical and Computational Semantics – Volume 1: Proceedings of the main conference and the s…
  15. Newton, and Mario E Occhialino Jr. 1993. Recognizing and enforcing state and tribal judgments: A roundtable discussion of law, policy, and practice. Am. Indian L. Rev., 18:239. find
  16. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. In International Conference on Learning Representations.
  17. Pritish Sahu, Michael Cogswell, Yunye Gong, and Ajay Divakaran. 2022. Unpacking large language models with conceptual consistency. arXiv preprint arXiv:2209.15093. — [arXiv:2209.15093](https://arxiv.org/abs/2209.15093) find
  18. Dinesh Babu Jayagopi, Hayley Hung, Chuohao Yeo, and Daniel Gatica-Perez. 2009. Modeling dominance in group conversations using nonverbal activity cues. IEEE Transactions on Audio, Speech, and Language Processing, 17(3):501–513. find
  19. Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, et al. 2022. Multitask prompted training enables zero-shot task generalization. In International Conference on Learning Representations.
  20. Jaehun Jung, Lianhui Qin, Sean Welleck, Faeze Brahman, Chandra Bhagavatula, Ronan Le Bras, and Yejin Choi. 2022. Maieutic prompting: Logically consistent reasoning with recursive explanations. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1266–1279,…
  21. Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. 2019. Social iqa: Commonsense reasoning about social interactions. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processi… find
  22. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. In Advances in Neural Information Processing Systems.
  23. Timo Schick, Jane Dwivedi-Yu, Zhengbao Jiang, Fabio Petroni, Patrick Lewis, Gautier Izacard, Qingfei You, Christoforos Nalmpantis, Edouard Grave, and Sebastian Riedel. 2022. Peer: A collaborative language model. arXiv preprint arXiv:2208.11663. find
  24. Yifei Li, Zeqi Lin, Shizhuo Zhang, Qiang Fu, Bei Chen, Jian-Guang Lou, and Weizhu Chen. 2022. On the advance of making language models better reasoners. arXiv preprint arXiv:2206.02336. — [arXiv:2206.02336](https://arxiv.org/abs/2206.02336) find
  25. Igor Mayer. 1997. Debating technologies. A Methodological Contribution to the Design and Evaluation of Participatory Policy Analysis. Tilburg, The Netherlands. find
  26. OpenAI. 2023. Gpt-4 technical report.
  27. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730–27744.
  28. Joon Sung Park, Joseph C O’Brien, Carrie J Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442. find
  29. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485–5551.
  30. Richard E Ransom, Christine Zuni, Philip Sam Deloria, Robert N Clinton, Robert Laurence, Nell Jessup Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. Commonsenseqa: A question answering challenge targeting commonsense knowledge. In Proceedings of the 2019 Conference of the N…
  31. Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. 2022. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239.
  32. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.
  33. Boshi Wang, Xiang Deng, and Huan Sun. 2022a. Iteratively prompt pre-trained language models for chain of thought. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2714–2730. find
  34. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. 2022b. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.
  35. Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Atharva Naik, Arjun Ashok, Arut Selvan Dhanasekaran, Anjana Arunkumar, David Stap, et al. 2022c. Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tasks. In Proceedings of the 202… find
  36. Prompt 1 : Zero-shot Prompt of ChatGPT and ChatGPT-0301 on αNLI, CommonsenseQA, COPA, e-CARE, Social IQa, and PIQA find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/xiong-2023-examining-inter-consistency-large-language/