Evaluating the Logical Reasoning Ability of ChatGPT and GPT-4

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-ND 4.0 (arXiv)" does not allow redistributing the text itself. Original — arxiv.org/pdf/2304.03439

Summary

Are ChatGPT and GPT-4 really strong at logic-heavy language tasks? The authors test both models on two task families. The first is multiple-choice reading comprehension: LogiQA 2.0 (English and Chinese), ReClor and AR-LSAT. The second is natural language inference: ConTRoL, ConjNLI, HELP, MED and TaxiNLI, with MNLI as a conventional reference point. To check robustness, they also build a new out-of-distribution set from 2022 Chinese Civil Servant Exam questions. The models are compared with a fine-tuned RoBERTa baseline and human scores, using the API and smaller manual samples in the Chat UI.

Both GPT models beat RoBERTa on most benchmarks, and GPT-4 is usually best, particularly on established datasets such as ReClor. Accuracy drops sharply on newer or out-of-distribution data and on NLI, where GPT-4 also fails to output labels consistently. The prompt-formatted datasets are released as the LogiEval suite, a reusable yardstick for reasoning in prompt-based LLMs.

Key points

GPT-4 (API) accuracy versus RoBERTa / ChatGPT:
- ReClor dev: 87.20% vs 55.01 / 57.38
- LogiQA 2.0 test: 72.25% vs 48.76 / 52.37
- LogiQA 2.0 ood: 58.49% vs 33.22 / 38.44
- AR-LSAT: only 33.48%, with ChatGPT at 20.42%
On NLI, GPT-4 reaches 89.42% on MED and 72.71% on ConjNLI. It manages only 46.01% on HELP, 60.08% on TaxiNLI and 64.08% on MNLI dev.
Context matters: when 20 questions were asked in one conversation rather than in separate sessions, GPT-4 answered more of them correctly.
- LogiQA 2.0 ood: 9/20 vs 5/20
- ConTRoL: 13/20 vs 7/20
Of GPT-4's first 10 LogiQA mistakes, 4 were logical fallacies, 3 were scope errors and 3 came from unresolved semantic ambiguity.
The resources are public: LogiEval at github.com/csitfun/LogiEval and the out-of-distribution data in the LogiQA 2.0 repository.

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: Other

Cited by

112 papers of our library cite it (of the 140673 whose bibliographies we read); 9 of them have a page here, 103 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2304.03439 · arXiv 2304.03439

@misc{liu2023evaluating,
  title = {{Evaluating the Logical Reasoning Ability of ChatGPT and GPT-4}},
  author = {Liu, Hanmeng and Ning, Ruoxi and Teng, Zhiyang and Liu, Jian and Zhou, Qiji and Zhang, Yue},
  year = {2023},
  doi = {10.48550/arxiv.2304.03439},
  eprint = {2304.03439},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2304.03439}
}

References

27 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, Quyet V. Do, Yan Xu, and Pascale Fung. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. find our card
Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632–642, Lisbon, Portugal. Association for Computati… find
Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter,… find
Maxwell John Cresswell. 1973. Logics and languages (1st ed.). Routledge. find
Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In Proc. of ICML. find
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR. find
Lucja Iwańska. 1993. Logical reasoning in natural language: It is all about knowledge. Minds and Machines. find
Pratik Joshi, Somak Aditya, Aalok Sathe, and Monojit Choudhury. 2020. Taxinli: Taking a ride up the NLU hill. CoRR. find
Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2023. Large language models are zero-shot reasoners. find our card
Robert Kowalski. 1979. Logic for problem solving, volume 7. Ediciones Díaz de Santos. find
Tao Li and Vivek Srikumar. 2019. Augmenting neural networks with first-order logic. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 292–302, Florence, Italy. Association for Computational Linguistics. find
Hanmeng Liu, Leyang Cui, Jian Liu, and Yue Zhang. 2020. Natural language inference in context - investigating contextual reasoning over long texts. CoRR. find
Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V. Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv. find
Bill MacCartney and Christopher D. Manning. 2007a. Natural logic for textual inference. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing. find
Bill MacCartney and Christopher D Manning. 2007b. Natural logic for textual inference. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing. find
OpenAI. 2023. Gpt-4 technical report. find our card
Chengwei Qin, Aston Zhang, Zhuosheng Zhang, Jiaao Chen, Michihiro Yasunaga, and Diyi Yang. 2023. Is chatgpt a general-purpose natural language processing task solver? find our card
Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. CoRR. find
Swarnadeep Saha, Yixin Nie, and Mohit Bansal. 2020. Conjnli: Natural language inference over conjunctive sentences. CoRR. find
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. CoRR. find
Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. find
GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP. find our card
Siyuan Wang, Zhongkun Liu, Wanjun Zhong, Ming Zhou, Zhongyu Wei, Zhumin Chen, and Nan Duan. 2022. From lsat: The progress and challenges of complex reasoning. IEEE/ACM Transactions on Audio, Speech, and Language Processing. find
Adina Williams, Nikita Nangia, and Samuel Bowman. 2018a. A broad-coverage challenge corpus for sentence understanding through inference. In Proc. of NAACL. find
Adina Williams, Nikita Nangia, and Samuel Bowman. 2018b. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long P… find
Hitomi Yanaka, Koji Mineshima, Daisuke Bekki, Kentaro Inui, Satoshi Sekine, , and Johan Bos. 2019a. Help: A dataset for identifying shortcomings of neural models in monotonicity reasoning. In Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM2019). find
Hitomi Yanaka, Koji Mineshima, Daisuke Bekki, Kentaro Inui, Satoshi Sekine, Lasha Abzianidze, and Johan Bos. 2019b. Can neural networks understand monotonicity reasoning? find