Elamur Research⌕
☰
home / papers / 2023 / The ICL Consistency Test

The ICL Consistency Test

Paper · 2023 · Lucas Weber, Elia Bruni and Dieuwke Hupkes

Summary

Large language models adapted through in-context learning (ICL) often give different answers to the same input when the prompt changes in ways that should not matter, such as rewording the instruction. Accuracy alone does not reveal this. Two prompts can score equally well while getting different items right. The authors propose the ICL consistency test as a contribution to the GenBench Collaborative Benchmarking Task. It provides preprocessed few-shot prompts built from ANLI and MNLI items wrapped in promptsource (P3) templates. Each prompt follows one of 96 setups, defined by combining mostly binary prompt-design factors. Consistency is measured with Cohen's κ, both for each factor and as a model-level average. The authors tested eight models: 7b to 65b versions in vanilla and instruction-tuned (IT) forms, evaluated on ANLI. Accuracy varied enormously from setup to setup, and no model was consistent under any factor change. IT models were more stable, and their stability improved with size. The test shows that current LLMs generalise in a non-robust way, and it offers a complementary signal to accuracy.

Key points

  • Data: 600 validation items each from ANLI (sampled equally from its three rounds) and MNLI. Instructions come from P3 templates, and in-context examples are drawn from the training sets.
  • The factors are:
    • number of shots (2 vs. 5)
    • high- vs. low-performing instructions
    • balanced labels
    • cross-templates
    • cross-task examples from QQP
    • alternative equivalent instructions
    • single-label contexts
  • Accuracy for large models ranged from below chance up to 67%, depending on setup. For the largest model on ANLI, the gap reached about 40%.
  • Per-factor κ never exceeded 0.75. Models tended to be sensitive to the same factors (e.g., HP instructions) and comparatively robust to others (e.g., balanced labels).
  • Main limitations: the test covers only natural language inference, and its factor list is not exhaustive. Users can add their own factors.

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, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputational Theory and Mathematics · 252 worksDecision Sciences · 209 worksSoftware · 92 worksInformation Systems and Management · 41 works

Cite

Cite as Lucas Weber, Elia Bruni and Dieuwke Hupkes. 2023. The ICL Consistency Test. arXiv:2312.04945.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2312.04945 · arXiv 2312.04945

@misc{weber2023icl,
  title = {{The ICL Consistency Test}},
  author = {Lucas Weber, Elia Bruni and Dieuwke Hupkes},
  year = {2023},
  doi = {10.48550/arxiv.2312.04945},
  eprint = {2312.04945},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2312.04945}
}

References

29 references, 17 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. In The Eleventh International Conference on Learning Representations.
  2. Anas Awadalla, Mitchell Wortsman, Gabriel Ilharco, Sewon Min, Ian Magnusson, Hannaneh Hajishirzi, and Ludwig Schmidt. 2022. Exploring the landscape of distributional robustness for question answering models. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5971–5987, A… find
  3. Stephen Bach, Victor Sanh, Zheng Xin Yong, Albert Webson, Colin Raffel, Nihal V. Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, Zaid Alyafeai, Manan Dey, Andrea Santilli, Zhiqing Sun, Srulik Ben-david, Canwen Xu, Gunjan Chhablani, Han Wang, Jason Fries, Maged Al-shaibani, Shanya… find
  4. Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. 2023. A multi-task, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. ArXiv preprint, abs/2302.04023.
  5. 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,… — Language Models are Few-Shot Learners
  6. Shivam Garg, Dimitris Tsipras, Percy S Liang, and Gregory Valiant. 2022. What can transformers learn in-context? a case study of simple function classes. Advances in Neural Information Processing Systems, 35:30583–30598. find
  7. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.
  8. D. Hupkes, M. Giulianelli, V. Dankers, et al. 2023. A taxonomy and review of generalization research in nlp. Nature Machine Intelligence, 5:1161–1174. find
  9. Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, et al. 2022. Holistic evaluation of language models. ArXiv preprint, abs/2211.09110. find
  10. Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p…
  11. Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11048–11064, Ab…
  12. Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. 2022. Reframing instructional prompts to GPTk’s language. In Findings of the Association for Computational Linguistics: ACL 2022, pages 589–612, Dublin, Ireland. Association for Computational Linguistics.
  13. Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885–4901, Online. Association for Computa…
  14. Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37–46. find
  15. Xenia Ohmer, Elia Bruni, and Dieuwke Hupkes. 2023. Evaluating task understanding through multilingual consistency: A chatgpt case study. ArXiv preprint, abs/2305.11662. — [arXiv:2305.11662](https://arxiv.org/abs/2305.11662) find
  16. Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022. Llm. int8 (): 8-bit matrix multiplication for transformers at scale. ArXiv preprint, abs/2208.07339. find
  17. 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
  18. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392, Austin, Texas. Association for Computational Linguistics.
  19. Chenglei Si, Zhe Gan, Zhengyuan Yang, Shuohang Wang, Jianfeng Wang, Jordan Boyd-Graber, and Lijuan Wang. 2023. Prompting gpt-3 to be reliable.
  20. Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca. ↗
  21. 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, abs/2302.13971.
  22. Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2022. Self-instruct: Aligning language model with self generated instructions. ArXiv preprint, abs/2212.10560. find
  23. Zhiguo Wang, Wael Hamza, and Radu Florian. 2017. Bilateral multi-perspective matching for natural language sentences. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI 2017, Melbourne, Australia, August 19-25, 2017, pages 4144–4150. ijcai.org.
  24. Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022. Finetuned language models are zero-shot learners. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview…
  25. Jerry Wei, Jason Wei, Yi Tay, Dustin Tran, Albert Webson, Yifeng Lu, Xinyun Chen, Hanxiao Liu, Da Huang, Denny Zhou, et al. 2023. Larger language models do in-context learning differently. ArXiv preprint, abs/2303.03846.
  26. Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. 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 Pa…
  27. Kang Min Yoo, Junyeob Kim, Hyuhng Joon Kim, Hyunsoo Cho, Hwiyeol Jo, Sang-Woo Lee, Sang-goo Lee, and Taeuk Kim. 2022. Ground-truth labels matter: A deeper look into input-label demonstrations. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2422– 2437…
  28. Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Lea…
  29. Ruiqi Zhong, Kristy Lee, Zheng Zhang, and Dan Klein. 2021. Adapting language models for zero-shot learning by meta-tuning on dataset and prompt collections. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2856–2878, Punta Cana, Dominican Republic. Association for Comp… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/weber-2023-icl-consistency-test/