Elamur Research⌕
☰
home / papers / 2022 / Explanations from Large Language Models Make Small Reasoners Better

Explanations from Large Language Models Make Small Reasoners Better

Paper · 2022 · Shiyang Li, Jianshu Chen, Yelong Shen et al.

Summary

Small language models (SLM) are cheaper to store and run than LLMs, but they fall well behind on complex reasoning. Earlier work found that fine-tuning T5 with crowdsourced human explanations brought no real gains, possibly because those explanations are often noisy. This paper tests whether explanations generated automatically by GPT-3 make a better training signal. Starting from only seven hand-written demonstrations, the authors prompt GPT-3 to explain every training example in three ways:

  • COTE: chain of thought prompting, keeping an explanation only when the LLM's answer is correct.
  • Rationalization: explanations conditioned on the gold label.
  • Hybrid: chain of thought where it succeeds, rationalization otherwise.

T5 is then fine-tuned in a multi-task setup. This covers the existing MT-Re and MT-Ra variants and a new MT-CoT, which learns two tasks: answering directly, and writing an explanation before the answer. Every combination improved on plain single-task fine-tuning across three reasoning datasets. On CommonsenseQA the small model even surpassed GPT-3 175B. The result suggests that cheap, deployable reasoners can gain both accuracy and explainability from LLM-written rationales.

Key points

  • Gains over standard T5 fine-tuning reach 8.1% accuracy. On CommonsenseQA the method beats fine-tuned or prompted GPT-3 (175B, about 60x larger) by up to 9.5%.
  • The hybrid explanation strategy was best on two of the three datasets. MT-CoT also beat MT-Re and MT-Ra on two of three.
  • Unlike Hase et al. (2020), the same MT-Re/MT-Ra setups give clear improvements once the explanations come from an LLM instead of crowdworkers.
  • In the human evaluation, three annotators compared T5-3B and GPT-3 explanations on 100 correctly answered CommonsenseQA items:
    • T5 won 14% of comparisons, 44% were ties, and GPT-3 won 42%.
    • So 58% of T5's explanations were competitive with or better than GPT-3's.
  • Limitations: the method needs a tuned task weight α and covers only textual reasoning. Biases from the LLM may also carry over into the small model, though explanation output can be switched off.

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

🏢MicrosoftRedmond · United States378 works · 161,630 citations🏢TencentShenzhen · China163 works · 21,032 citations🎓University of WaterlooWaterloo · Canada80 works · 15,030 citations🎓University of California, Santa BarbaraSanta Barbara · United States67 works · 10,550 citations🔬Vector InstituteToronto · Canada37 works · 5,703 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,894 worksReasoning · 1,685 works

Cite

Cite as Shiyang Li, Jianshu Chen, Yelong Shen, Zhiyu Chen, Xinlu Zhang, Zekun Li, Hong Wang, Jing Qian, Baolin Peng, Yi Mao, Wenhu Chen and Xifeng Yan. 2022. Explanations from Large Language Models Make Small Reasoners Better. arXiv:2210.06726.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2210.06726 · arXiv 2210.06726

@misc{li2022explanations,
  title = {{Explanations from Large Language Models Make Small Reasoners Better}},
  author = {Shiyang Li, Jianshu Chen, Yelong Shen, Zhiyu Chen, Xinlu Zhang, Zekun Li, Hong Wang, Jing Qian, Baolin Peng, Yi Mao, Wenhu Chen and Xifeng Yan},
  year = {2022},
  doi = {10.48550/arxiv.2210.06726},
  eprint = {2210.06726},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2210.06726}
}

References

40 references, 25 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abubakar Abid, Maheen Saleem Farooqi, and James Y. Zou. 2021. Persistent anti-muslim bias in large language models. Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society.
  2. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D 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 Ziegler, Jeffrey Wu, Clemens Winter, Chri… — Language Models are Few-Shot Learners
  3. Oana-Maria Camburu, Tim Rocktäschel, Thomas Lukasiewicz, and Phil Blunsom. 2018. e-snli: Natural language inference with natural language explanations. In NeurIPS.
  4. Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek B Rao, Parker Barnes, Yi Tay, Noam M. Shazeer, Vinodkumar Prabhakaran… — PaLM: Scaling Language Modeling with Pathways
  5. Miruna-Adriana Clinciu, Arash Eshghi, and Helen Hastie. 2021. A study of automatic metrics for the evaluation of natural language explanations. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2376–2387, Online. Associ… find
  6. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. ArXiv, abs/2110.14168.
  7. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,…
  8. 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
  9. Peter Hase, Shiyue Zhang, Harry Xie, and Mohit Bansal. 2020. Leakage-adjusted simulatability: Can models generate non-trivial explanations of their behavior in natural language? In EMNLP (Findings), pages 4351–4367.
  10. Lisa Anne Hendricks, Zeynep Akata, Marcus Rohrbach, Jeff Donahue, Bernt Schiele, and Trevor Darrell. 2016. Generating visual explanations. In ECCV. find
  11. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780. find
  12. Sarthak Jain, Sarah Wiegreffe, Yuval Pinter, and Byron C. Wallace. 2020. Learning to faithfully rationalize by construction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4459–4473, Online. Association for Computational Linguistics.
  13. W. Lewis Johnson. 1994. Agents that learn to explain themselves. AAAI. find
  14. Maxime Kayser, Oana-Maria Camburu, Leonard Salewski, Cornelius Emde, Virginie Do, Zeynep Akata, and Thomas Lukasiewicz. 2021. e-vil: A dataset and benchmark for natural language explanations in vision-language tasks. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1224–1234. find
  15. Pang Wei Koh and Percy Liang. 2017. Understanding black-box predictions via influence functions. ArXiv, abs/1703.04730.
  16. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. ArXiv, abs/2205.11916.
  17. Andrew Kyle Lampinen, Ishita Dasgupta, Stephanie C. Y. Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L. McClelland, Jane X. Wang, and Felix Hill. 2022. Can language models learn from explanations in context? ArXiv, abs/2204.02329.
  18. Shiyang Li, Jianshu Chen, and Dian Yu. 2019. Teaching pretrained models with commonsense reasoning: A preliminary kb-based approach. ArXiv, abs/1909.09743. — [arXiv:1909.09743](https://arxiv.org/abs/1909.09743) find
  19. Li Lucy and David Bamman. 2021. Gender and representation bias in GPT-3 generated stories. In Proceedings of the Third Workshop on Narrative Understanding, pages 48–55, Virtual. Association for Computational Linguistics.
  20. Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct electricity? a new dataset for open book question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2381–2391, Brussels, Belgium. Association fo… find
  21. Sharan Narang, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel, and Karishma Malkan. 2020. Wt5?! training text-to-text models to explain their predictions. ArXiv, abs/2004.14546.
  22. Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training. https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/language-unsupervised/ language understanding paper.pdf. ↗
  23. Alec Radford, Jeffrey Wu, Rewon Child, Luan David, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI Blog. find
  24. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, W. Li, and Peter J. Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. ArXiv, abs/1910.10683.
  25. Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. Explain yourself! leveraging language models for commonsense reasoning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4932–4942, Florence, Italy. Association for Computat…
  26. Wojciech Samek, Grégoire Montavon, Andrea Vedaldi, Lars Kai Hansen, and Klaus-Robert Müller. 2019. Explainable AI: interpreting, explaining and visualizing deep learning, volume 11700. Springer Nature. find
  27. 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 North American Chapter of the Association for Computational Linguistics: Human Language Technologies,…
  28. Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam M. Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, Yaguang Li, Hongrae Lee, Huaixiu Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong X… — LaMDA: Language Models for Dialog Applications
  29. Ben Wang. 2021. Mesh-Transformer-JAX: Model-Parallel Implementation of Transformer Language Model with JAX. https://github.com/kingoflolz/mesh-transformer-jax. ↗
  30. Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V Le. 2022a. Finetuned language models are zero-shot learners. In International Conference on Learning Representations.
  31. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022b. Chain of thought prompting elicits reasoning in large language models. ArXiv, abs/2201.11903.
  32. Sarah Wiegreffe, Jack Hessel, Swabha Swayamdipta, Mark O. Riedl, and Yejin Choi. 2021a. Reframing human-ai collaboration for generating free-text explanations. ArXiv, abs/2112.08674.
  33. Sarah Wiegreffe and Ana Marasović. 2021. Teach me to explain: A review of datasets for explainable natural language processing. In NeurIPS Datasets and Benchmarks. find
  34. Sarah Wiegreffe, Ana Marasović, and Noah A. Smith. 2021b. Measuring association between labels and free-text rationales. In EMNLP.
  35. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier-ric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lho… — Transformers: State-of-the-Art Natural Language Processing
  36. Yichong Xu, Chenguang Zhu, Shuohang Wang, Siqi Sun, Hao Cheng, Xiaodong Liu, Jianfeng Gao, Pengcheng He, Michael Zeng, and Xuedong Huang. 2021. Human parity on commonsenseqa: Augmenting self-attention with external attention. ArXiv, abs/2112.03254. — [arXiv:2112.03254](https://arxiv.org/abs/2112.03254) find
  37. Xi Ye and Greg Durrett. 2022. The unreliability of explanations in few-shot in-context learning. ArXiv, abs/2205.03401.
  38. E. Zelikman, Yuhuai Wu, and Noah D. Goodman. 2022. Star: Bootstrapping reasoning with reasoning. ArXiv, abs/2203.14465.
  39. Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Olivier Bousquet, Quoc Le, and Ed Chi. 2022. find
  40. Least-to-most prompting enables complex reasoning in large language models. ArXiv, abs/2205.10625.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/li-2022-explanations-large-language-models-make/