Elamur Research
☰
home / papers / 2023 / Large Language Models are Better Reasoners with Self-Verification

Large Language Models are Better Reasoners with Self-Verification

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/2023.findings-emnlp.167.pdf

Summary

Chain-of-thought (CoT) prompting helps large language models reason through multi-step problems. One wrong step, though, can carry through the chain and spoil the final answer. Earlier fixes trained a separate verifier model, which needs human labels and fine-tuning and gives scores that are hard to interpret. The authors show that an LLM can check its own answers with only a few prompts. First, in forward reasoning, the model samples several candidate answers with CoT. Then, in backward verification, each candidate is added to the problem as a known fact and one of the original conditions is hidden. The model tries to work that condition out again, and a candidate scores higher the more often the recovered value matches the true one. The highest-scoring candidate becomes the final answer. For arithmetic tasks they mask numeric conditions; for other tasks they turn the check into a true/false question. This raised accuracy on arithmetic, commonsense and logical reasoning benchmarks without any extra training. It also gives a verification signal that can be traced step by step.

Key points

  • Two verification variants: Condition Mask Verification for arithmetic tasks and True-False Item Verification for general tasks such as commonsense and logical reasoning.
  • Reported gains include GSM8K rising from 60.8 to 65.1 and SingleEq from 91.01 to 93.40. The method also improves on self-consistency and Least-to-Most prompting.
  • On InstructGPT cases with one correct candidate answer among wrong ones, verification picked the right answer more often than plain CoT (decoded at temperature 0) or a random guess. On GSM8K this was 58.9% vs 53.4% vs 35.7%, with the largest gains on MultiArith and SingleEq.
  • Masking several conditions usually beat masking just one. Even P = 2 verification samples improved on the CoT baseline, and the authors suggest P ≈ 10 as a balance between cost and accuracy.
  • Limitations: it only works if a correct answer is among the candidates, it helps strong models more than small ones, it costs extra compute, and it checks final conclusions rather than the reasoning steps. Code: github.com/WENGSYX/Self-Verification.

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: Mathematics & statistics

Cited by

195 papers of our library cite it (of the 140673 whose bibliographies we read); 13 of them have a page here, 182 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2023.findings-emnlp.167 · arXiv 2212.09561

@inproceedings{weng2023large,
  title = {{Large Language Models are Better Reasoners with Self-Verification}},
  author = {Weng, Yixuan and Zhu, Minjun and Xia, Fei and Li, Bin and He, Shizhu and Liu, Shengping and Sun, Bin and Liu, Kang and Zhao, Jun},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2023},
  year = {2023},
  doi = {10.18653/v1/2023.findings-emnlp.167},
  eprint = {2212.09561},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/2023.findings-emnlp.167}
}

References

68 references, 26 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ekin Akyürek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. 2022. What learning algorithm is in-context learning? investigations with linear models. arXiv preprint arXiv:2211.15661. find our card
  2. Aida Amini, Saadia Gabriel, Peter Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019. Mathqa: Towards interpretable math word problem solving with operation-based formalisms. north american chapter of the association for computational linguistics. find our card
  3. Patel Arkil, Bhattamishra Satwik, and Goyal Navin. 2021. Are nlp models really able to solve simple math word problems? find our card
  4. Simran Arora, Avanika Narayan, Mayee F. Chen, Laurel J. Orr, Neel Guha, Kush Bhatia, Ines Chami, Frederic Sala, and Christopher Ré. 2022. Ask me anything: A simple strategy for prompting language models. find
  5. Akari Asai and Hannaneh Hajishirzi. 2020. Logic-guided data augmentation and regularization for consistent question answering. meeting of the association for computational linguistics. find
  6. Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman, Hannah Rashkin, Doug Downey, Scott Wen tau Yih, and Yejin Choi. 2019. Abductive commonsense reasoning. Learning. find our card
  7. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901. find
  8. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harrison Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mik… find our card
  9. Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang, Weihua Peng, Ming Liu, Bing Qin, and Ting Liu. 2023. A survey of chain of thought reasoning: Advances, frontiers and future. find
  10. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168. find our card
  11. Xiang Deng, Yu Su, Alyssa Lees, You Wu, Cong Yu, and Huan Sun. 2021. Reasonbert: Pre-trained to reason with distant supervision. empirical methods in natural language processing. find
  12. Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2023. Chain-of-verification reduces hallucination in large language models. ArXiv, abs/2309.11495. find our card
  13. Paul Erd. 1970. On a new law of large numbers. J. Anal. Math, 23(103111):8. find
  14. Yao Fu, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. 2022. Complexity-based prompting for multi-step reasoning. find our card
  15. Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Pal: Program-aided language models. find
  16. 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. find
  17. Yuxian Gu, Jiaxin Wen, Hao Sun, Yi Song, Pei Ke, Chujie Zheng, Zheng Zhang, Jianzhu Yao, Lei Liu, Xiaoyan Zhu, and Minlie Huang. 2023. Eva2.0: Investigating open-domain chinese dialogue systems with large-scale pre-training. Machine Intelligence Research, 20(2):207–219. find
  18. Charles R Harris, K Jarrod Millman, Stéfan J Van Der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J Smith, et al. 2020. Array programming with numpy. Nature, 585(7825):357–362. find our card
  19. Namgyu Ho, Laura Schmid, and Se-Young Yun. 2022. Large language models are reasoning teachers. find our card
  20. Mohammad Javad Hosseini, Hannaneh Hajishirzi, Oren Etzioni, and Nate Kushman. 2014. Learning to solve arithmetic word problems with verb categorization. empirical methods in natural language processing. find our card
  21. Minghao Hu, Yuxing Peng, Zhen Huang, and Dongsheng Li. 2019. A multi-type multi-span network for reading comprehension that requires discrete reasoning. empirical methods in natural language processing. find
  22. Tianbo Ji, Chenyang Lyu, Gareth Jones, Liting Zhou, and Yvette Graham. 2022. Qascore – an unsupervised unreferenced metric for the question generation evaluation. find
  23. 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. find our card
  24. Rik Koncel-Kedziorski, Hannaneh Hajishirzi, Ashish Sabharwal, Oren Etzioni, and Siena Dumas Ang. 2015. Parsing algebraic word problems into equations. Transactions of the Association for Computational Linguistics. find
  25. Nate Kushman, Yoav Artzi, Luke Zettlemoyer, and Regina Barzilay. 2014. Learning to automatically solve algebra word problems. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 271–281, Baltimore, Maryland. Association for Comput… find
  26. Bin Li, Encheng Chen, Hongru Liu, Yixuan Weng, Bin Sun, Shutao Li, Yongping Bai, and Meiling Hu. 2021. More but correct: Generating diversified and entity-revised medical response. arXiv preprint arXiv:2108.01266. — [arXiv:2108.01266](https://arxiv.org/abs/2108.01266) find
  27. Jierui Li, Lei Wang, Jipeng Zhang, Yan Wang, Bing Tian Dai, and Dongxiang Zhang. 2019. Modeling intra-relation in math word problems with different functional multi-head attentions. meeting of the association for computational linguistics. find
  28. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training… find
  29. Arkil Patel, Satwik Bhattamishra, and Navin Goyal. 2021. Are nlp models really able to solve simple math word problems. north american chapter of the association for computational linguistics. find our card
  30. Xinyu Pi, Qian Liu, Bei Chen, Morteza Ziyadi, Zeqi Lin, Yan Gao, Qiang Fu, Jian-Guang Lou, and Weizhu Chen. 2022. Reasoning like program executors. find
  31. David L Poole and Alan K Mackworth. 2010. Artificial Intelligence: foundations of computational agents. Cambridge University Press. find
  32. Shuofei Qiao, Yixin Ou, Ningyu Zhang, Xiang Chen, Yunzhi Yao, Shumin Deng, Chuanqi Tan, Fei Huang, and Huajun Chen. 2022. Reasoning with language model prompting: A survey. find our card
  33. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. find
  34. Jinghui Lu, Rui Zhao, Brian Mac Namee, Dongsheng Zhu, Weidong Han, and Fei Tan. 2022. What makes pre-trained language models better zero/few-shot learners? find
  35. Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, H. Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks, Maribeth Rauh, Po-S… find
  36. Aman Madaan, Shuyan Zhou, Uri Alon, Yiming Yang, and Graham Neubig. 2022. Language models of code are few-shot commonsense learners. find
  37. Subhro Roy and Dan Roth. 2016. Solving general arithmetic word problems. arXiv: Computation and Language. find our card
  38. 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
  39. Hunter Lightman, Vineet Kosaraju, Yura Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let’s verify step by step. ArXiv, abs/2305.20050. find our card
  40. Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunsom. 2017. Program induction by rationale generation : Learning to solve and explain algebraic word problems. meeting of the association for computational linguistics. find
  41. Jian Liu, Leyang Cui, Hanmeng Liu, Dandan Huang, Yile Wang, and Yue Zhang. 2020. Logiqa: A challenge dataset for machine reading comprehension with logical reasoning. international joint conference on artificial intelligence. find our card
  42. Abulhair Saparov and He He. 2022. Language models are greedy reasoners: A systematic formal analysis of chain-of-thought. find our card
  43. Jianhao Shen, Yichun Yin, Lin Li, Lifeng Shang, Xin Jiang, Ming Zhang, and Qun Liu. 2021. Generate & rank: A multi-task framework for math word problems. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2269–2279. find
  44. Yibin Shen and Cheqing Jin. 2020. Solving math word problems with multi-encoders and multi-decoders. international conference on computational linguistics. find
  45. Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, Agnieszka Kluska, Aitor Lewkowycz, Akshat Agarwal, Alethea Power, Alex Ray, Alex Warstadt, Alexander W. Kocurek, Ali Safaya, Ali Tazarv, Al… find our card
  46. Tian-Xiang Sun, Xiang-Yang Liu, Xi-Peng Qiu, and Xuan-Jing Huang. 2022. Paradigm shift in natural language processing. Machine Intelligence Research, 19(3):169–183. find
  47. Xingwu Sun, Jing Liu, Yajuan Lyu, Wei He, Yanjun Ma, and Shi Wang. 2018. Answer-focused and position-aware neural question generation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3930–3939, Brussels, Belgium. Association for Computational Linguist… find
  48. Siyuan Wang, Wanjun Zhong, Duyu Tang, Zhongyu Wei, Zhihao Fan, Daxin Jiang, Ming Zhou, and Nan Duan. 2022b. Logic-driven context extension and data augmentation for logical reasoning of text. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1619–1629, Dublin, Ireland. As… find
  49. Xinyi Wang, Wanrong Zhu, Michael Saxon, Mark Steyvers, and William Yang Wang. 2023a. Large language models are implicitly topic models: Explaining and finding good demonstrations for in-context learning. find
  50. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023b. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations. find our card
  51. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023c. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations. find our card
  52. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903. find our card
  53. Yixuan Weng, Bin Li, Fei Xia, Minjun Zhu, Bing Sun, Shizhu He, Kang Liu, and Jun Zhao. 2023a. Large language models need holistically thought in medical conversational qa. ArXiv, abs/2305.05410. — [arXiv:2305.05410](https://arxiv.org/abs/2305.05410) find
  54. Yixuan Weng, Minjun Zhu, Fei Xia, Bin Li, Shizhu He, Kang Liu, and Jun Zhao. 2023b. Mastering symbolic operations: Augmenting language models with compiled neural networks. find
  55. Fei Xia, Bin Li, Yixuan Weng, Shizhu He, Kang Liu, Bin Sun, Shutao Li, and Jun Zhao. 2022. Medconqa: Medical conversational question answering system based on knowledge graphs. In Conference on Empirical Methods in Natural Language Processing. find
  56. Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2018. Commonsenseqa: A question answering challenge targeting commonsense knowledge. north american chapter of the association for computational linguistics. find
  57. Yisheng Xiao, Lijun Wu, Junliang Guo, Juntao Li, Min Zhang, Tao Qin, and Tie-yan Liu. 2022. A survey on non-autoregressive generation for neural machine translation and beyond. find
  58. Boshi Wang, Sewon Min, Xiang Deng, Jiaming Shen, You Wu, Luke Zettlemoyer, and Huan Sun. 2022a. Towards understanding chain-of-thought prompting: An empirical study of what matters. In Annual Meeting of the Association for Computational Linguistics. find our card
  59. Yichong Xu, Chenguang Zhu, Shuohang Wang, Siqi Sun, Hao Cheng, Xiaodong Liu, Jianfeng Gao, Pengcheng He, Michael Zeng, and Xuedong Huang. 2022. Human parity on commonsenseqa: Augmenting self-attention with external attention. find
  60. Lei Wang, Yan Wang, Deng Cai, Dongxiang Zhang, and Xiaojiang Liu. 2018. Translating a math word problem to a expression tree. empirical methods in natural language processing. find
  61. Ori Yoran, Alon Talmor, and Jonathan Berant. 2021. Turning tables: Generating examples from semi-structured tables for endowing language models with reasoning skills. arXiv: Computation and Language. find
  62. GSM8K: https://github.com/openai/ grade-school-math Weihao Yu, Zihang Jiang, Yanfei Dong, and Jiashi Feng. 2020. Reclor: A reading comprehension dataset requiring logical reasoning. international conference on learning representations. https://gitlab.cs. • SingleEq: washington.edu/ALGES/TACL2015 ↗ our card
  63. Zihan Yu, Liang He, Zhen Wu, Xinyu Dai, and Jiajun Chen. 2023. Towards better chain-of-thought prompting strategies: A survey. find
  64. AddSub: https://www.cs.washington. edu/nlp/arithmetic Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. 2022. Automatic chain of thought prompting in large language models. ↗ our card
  65. MultiArith: http://cogcomp.cs. illinois.edu/page/resource\_view/98 ↗
  66. Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, and Ed Chi. 2023. Least-to-most prompting enables complex reasoning in large language models. find our card
  67. AQUA-RAT: deepmind/AQuA • SVAMP: https://github.com/ arkilpatel/SVAMP Minjun Zhu, Bin Li, Yixuan Weng, and F. Xia. 2022a. A knowledge storage and semantic space alignment method for multi-documents dialogue generation. In Workshop on Document-grounded Dialogue and Conversational Question Answering.… ↗
  68. Minjun Zhu, Yixuan Weng, Shizhu He, Cunguang Wang, Kang Liu, Li Cai, and Jun Zhao. 2023. Learning to build reasoning chains by reliable path retrieval. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE. find