Shortcut Learning of Large Language Models in Natural Language Understanding
Open accessThe DOI matches its Crossref record; 41 of the 41 references it cites were found in a registry or an index.
Abstract
Shortcuts often hinder the robustness of large language models.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“Recent works show that even Large-Language Models (LLMs), pre-trained on text amounting to billions of words, suffer from shortcut learning, largely due to collection artifacts in training data [7], [21].”
From Language-Based Augmentation to Address Shortcut Learning in Object-Goal Navigation (2023) · cites this paper as [21]
Citations
Cites
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372 · cited 12×
“Recently, large language models (LLMs), such as BERT [8], RoBERTa [20], T5 [30], GPT-3 [4], have been reported to achieve state-of-the-art performance in a series of high-level NLU tasks.”
BackgroundInfluentialAdvances in neural information processing systems 7
Cios, K., & Shields, M. (1997) · Neurocomputing · doi:10.1016/s0925-2312(97)90018-7 · Cited by 22,059
Advances in neural information processing systems 7
(1996) · Computers & Mathematics with Applications · doi:10.1016/s0898-1221(96)90463-0 · Cited by 14,281
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929 · cited 1×
“Natural language understanding (NLU) is a subfield of artificial intelligence that requires computer software to comprehend input in the form of sentences.”
BackgroundChain-Of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q. V., & Zhou, D. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-1800 · Cited by 4,505
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022) · arXiv · doi:10.48550/arxiv.2201.11903 · Cited by 4,263 · cited 1×
“Note that this type of method can only be used when prior knowledge is known in advance about shortcuts.”
MethodRight for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference
McCoy, R. T., Pavlick, E., & Linzen, T. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1334 · Cited by 944 · cited 2×
“NLI, QA, and reading comprehension (Zhou and Bansal, 2020; McCoy et al., 2019).”
Arjovsky, M., Bottou, L., Gulrajani, I., & Lopez-Paz, D. (2019) · arXiv · doi:10.48550/arxiv.1907.02893 · Cited by 927 · cited 3×
“This is distinct from the literature that works with the toy task (e.g., Colored MNIST [1]), which typically contains a single type of bias and the bias is fully known.”
BackgroundMethodInfluentialJin, D., Jin, Z., Zhou, J. T., & Szolovits, P. (2020) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v34i05.6311 · Cited by 918 · cited 1×
“Besides, the adversarial examples created through TextFooler [15] are used to test the generalization of common sense reasoning models [3].”
MethodProbing Neural Network Comprehension of Natural Language Arguments
Niven, T., & Kao, H.-Y. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1459 · Cited by 395 · cited 3×
“on spurious statistical cues such as unigrams ‘not’, ‘do’, ‘is’ and bigrams ‘will not’ (Niven and Kao, 2019; Gururangan et al., 2018).”
BackgroundInfluentialDon’t Take the Easy Way Out: Ensemble Based Methods for Avoiding Known Dataset Biases
Clark, C., Yatskar, M., & Zettlemoyer, L. (2019) · Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) · doi:10.18653/v1/d19-1418 · Cited by 381 · cited 1×
“The goal is to train a debiased model by training it as an ensemble with a bias-only model (Clark et al., 2019; He et al., 2019; Sanh et al., 2021).”
MethodGradient Matching for Domain Generalization
Shi, Y., Seely, J., Torr, P. H. S., Siddharth, N., Hannun, A., Usunier, N., & Synnaeve, G. (2021) · Oxford University Research Archive (ORA) (University of Oxford) · doi:10.48550/arxiv.2104.09937 · Cited by 308 · cited 1×
“maximizing the inner product between gradients from different environments, with the goal of increasing model generalization (Shi et al., 2022).”
BackgroundSinha, K., Jia, R., Hupkes, D., Pineau, J., Williams, A., & Kiela, D. (2021) · Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/2021.emnlp-main.230 · Cited by 181 · cited 1×
“Recent ablation results indicate that word order does not matter for pre-trained language models [40].”
BackgroundTowards Debiasing Fact Verification Models
Schuster, T., Shah, D., Yeo, Y. J. S., Roberto Filizzola Ortiz, D., Santus, E., & Barzilay, R. (2019) · Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) · doi:10.18653/v1/d19-1341 · Cited by 171 · cited 2×
“Similarly, for the fact verification task, a symmetric test set is constructed that shares a philosophy similar to HANS [34].”
BackgroundShow Your Work: Scratchpads for Intermediate Computation with Language Models
Nye, M., Andreassen, A. J., Gur-Ari, G., Michalewski, H., Austin, J., Bieber, D., Dohan, D., Lewkowycz, A., Bosma, M., Luan, D., et al. (2021) · arXiv · doi:10.48550/arxiv.2112.00114 · Cited by 144 · cited 1×
“Note that this type of method can only be used when prior knowledge is known in advance about shortcuts.”
MethodGeneralization in NLI: Ways (Not) To Go Beyond Simple Heuristics
Bhargava, P., Валентинович, Д. А., & Rogers, A. (2021) · Workshop on Insights from Negative Results in NLP (Insights) · doi:10.18653/v1/2021.insights-1.18 · Cited by 119
Explaining Black Box Predictions and Unveiling Data Artifacts through Influence Functions
Han, X., Wallace, B. C., & Tsvetkov, Y. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.492 · Cited by 104
Towards Debiasing NLU Models from Unknown Biases
Utama, P. A., Moosavi, N. S., & Gurevych, I. (2020) · Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2020.emnlp-main.613 · Cited by 98
Incorporating Priors with Feature Attribution on Text Classification
Liu, F., & Avci, B. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1631 · Cited by 91 · cited 2×
“attribution explanations with rationale annotations created by domain experts, to enforce the model to make the right predictions for the right reasons (Liu and Avci, 2019).”
BackgroundAdversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models
Wang, B., Xu, C., Wang, S., Gan, Z., Cheng, Y., Gao, J., Awadallah, A. H., & Li, B. (2021) · arXiv · doi:10.48550/arxiv.2111.02840 · Cited by 88 · cited 1×
“Second, shortcut learning also results in models that could be easily fooled by adversarial samples, when small and often imperceptible human-crafted perturbations are added to the normal input (Wang et al., 2021a).”
BackgroundTableFormer: Robust Transformer Modeling for Table-Text Encoding
Yang, J., Gupta, A., Upadhyay, S., He, L., Goel, R., & Paul, S. (2022) · Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2022.acl-long.40 · Cited by 85
Pham, T., Bui, T., Mai, L., & Nguyen, A. (2021) · Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 · doi:10.18653/v1/2021.findings-acl.98 · Cited by 79
C2L: Causally Contrastive Learning for Robust Text Classification
Choi, S. L., Jeong, M., Han, H., & Hwang, S.-W. (2022) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v36i10.21296 · Cited by 68 · cited 1×
“A representative work presents a framework for mitigating spurious correlations using contrastive learning [6].”
BackgroundAvoiding the Hypothesis-Only Bias in Natural Language Inference via Ensemble Adversarial Training
Stacey, J., Minervini, P., Dubossarsky, H., Riedel, S., & Rocktäschel, T. (2020) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2020.emnlp-main.665 · Cited by 65 · cited 2×
“First, the task classifier and adversarial classifier jointly share the same encoder [42].”
BackgroundMethodWhat does BERT Learn from Multiple-Choice Reading Comprehension Datasets?Preprint
Si, C., Wang, S., Kan, M., & Jiang, J. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1910.12391 · Cited by 58
Combining Feature and Instance Attribution to Detect Artifacts
Pezeshkpour, P., Jain, S., Singh, S., & Wallace, B. (2022) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2022.findings-acl.153 · Cited by 52
Look at the First Sentence: Position Bias in Question Answering
Ko, M., Lee, J., Kim, H., Kim, G., & Kang, J. (2020) · Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2020.emnlp-main.84 · Cited by 50
MATE-KD: Masked Adversarial TExt, a Companion to Knowledge Distillation
Rashid, A., Lioutas, V., & Rezagholizadeh, M. (2021) · doi:10.18653/v1/2021.acl-long.86 · Cited by 49 · cited 2×
“It is typically implemented in two ways in the NLP domain [31, 42].”
BackgroundMethodRobustness Challenges in Model Distillation and Pruning for Natural Language Understanding
Du, M., Mukherjee, S., Cheng, Y., Shokouhi, M., Xia, H., & Awadallah, A. H. (2023) · Conference of the European Chapter of the Association for Computational Linguistics · doi:10.18653/v1/2023.eacl-main.129 · Cited by 47 · cited 1×
“Models have learned to rely on the erroneous text style as a shortcut rather than capturing the underlying semantics.”
BackgroundSelf-citationUnshuffling Data for Improved Generalization
Teney, D., Abbasnejad, E., & van den Hengel, A. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2002.11894 · Cited by 35
Avoiding Inference Heuristics in Few-shot Prompt-based Finetuning
Utama, P. A., Moosavi, N. S., Sanh, V., & Gurevych, I. (2021) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2021.emnlp-main.713 · Cited by 35 · cited 1×
“Preliminary research has been conducted to examine the shortcut learning challenge in the few-shot prompt-based fine-tuning paradigm [48].”
BackgroundGradient Methods Provably Converge to Non-Robust Networks
Vardi, G., Yehudai, G., & Shamir, O. (2022) · neural information processing systems · doi:10.52202/068431-1521 · Cited by 34
Why Machine Reading Comprehension Models Learn Shortcuts?
Lai, Y., Zhang, C., Feng, Y., Huang, Q., & Zhao, D. (2021) · Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 · doi:10.18653/v1/2021.findings-acl.85 · Cited by 32
Learning to Remove: Towards Isotropic Pre-trained BERT Embedding
Liang, Y., Cao, R., Zheng, J., Ren, J., & Gao, L. C. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-86383-8_36 · Cited by 30
To what extent do human explanations of model behavior align with actual model behavior?
Prasad, G., Nie, Y., Bansal, M., Jia, R., Kiela, D., & Williams, A. (2021) · BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP) · doi:10.18653/v1/2021.blackboxnlp-1.1 · Cited by 27
Towards Interpreting and Mitigating Shortcut Learning Behavior of NLU ModelsPreprint
Du, M., Manjunatha, V., Jain, R., Deshpande, R., Dernoncourt, F., Gu, J., Sun, T., & Hu, X. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2103.06922 · Cited by 11
Self-citationSimplicity Bias in 1-Hidden Layer Neural Networks
Morwani, D., Batra, J., Jain, P., & Netrapalli, P. (2023) · neural information processing systems · doi:10.52202/075280-0353 · Cited by 9
Gradient Methods Provably Converge to Non-Robust NetworksPreprint
Vardi, G., Yehudai, G., & Shamir, O. (2022) · arXiv (Cornell University) · doi:10.48550/arxiv.2202.04347 · Cited by 6 · cited 2×
“Another theoretical study has investigated the reason behind superficial correlations from the optimization perspective [49].”
BackgroundSimplicity Bias in 1-Hidden Layer Neural NetworksPreprint
Morwani, D., Batra, J., Jain, P., et al. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2302.00457 · Cited by 1 · cited 1×
“Note that this type of method can only be used when prior knowledge is known in advance about shortcuts.”
Method
Research works printed but not linked: 14
- Branco, R., Branco, A., Silva, J., and Rodrigues, J. Shortcutted commonsense: Data spuriousness in deep learning of commonsense reasoning. In Proceedings of the 2021 Conf. Empirical Methods in Natural Language Processing.
- Devlin, J., Chang, M., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. North American Chapter of the Assoc. Computational Linguistics, 2019.
- Ethayarajh, K. How contextual are contextualized word representations? Comparing the geometry of BERT, ELMo, and GPT-2 embeddings. In Proceedings of the 2019 Conf. Empirical Methods in Natural Language Processing and the 9th Intern. Joint Conf. Natural Language Processing, 55--65.
- Gururangan, S., Swayamdipta, S., Levy, O., Schwartz, R., Bowman, S.R., and Smith, N.A. Annotation artifacts in natural language inference data. North American Chapter of the Assoc. Computational Linguistics, 2018.
- Liu, Y. et al. RoBERTa: A robustly optimized BERT pretraining approach, 2019; arXiv:1907.11692. · arXiv
- Robinson, J., Sun, L., Yu, K., Batmanghelich, K., Jegelka, S., and Sra, S. Can contrastive learning avoid shortcut solutions? Advances in Neural Information Processing Systems, 2021.
- Sen, P. and Saffari, A. What do models learn from question answering datasets? In Proceedings of the 2020 Conf. Empirical Methods in Natural Language Processing.
- Si, C. et al. Prompting gpt-3 to be reliable, 2022; arXiv:2210.09150. · arXiv
- Stacey, J., Belinkov, Y., and Rei, M. Supervising Model Attention with Human Explanations for Robust Natural Language Inference. In Proceedings of the 2022 AAAI Conf. Artificial Intelligence.
- Sundararajan, M., Taly, A., and Yan, Q. Axiomatic attribution for deep networks. In Proceedings of the 2017 Intern. Conf. Machine Learning.
- Tu, L., Lalwani, G., Gella, S., and He, H. An empirical study on robustness to spurious correlations using pre-trained language models. Trans. Assoc. Computational Linguistics (2020).
- Webson, A. and Pavlick, E. Do prompt-based models really understand the meaning of their prompts? In Proceedings of the 2022 Conf. North American Chapter of the Assoc. Computational Linguistics: Human Language Technologies.
- Zellers, R., Bisk, Y., Schwartz, R., and Choi, Y. Swag: A large-scale adversarial dataset for grounded commonsense inference. In Proceedings of the 2018 Conf. Empirical Methods in Natural Language Processing.
- Zhao, Z., Wallace, E., Feng, S., Klein, D., and Singh, S. Calibrate before use: Improving few-shot performance of language models. In Proceedings of the 2021 Intern. Conf. Machine Learning, 12697--12706.
Cited by
The rise and potential of large language model based agents: a survey
Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al. (2025) · Science China Information Sciences · doi:10.1007/s11432-024-4222-0 · Cited by 970
“Furthermore, employing debiasing methods and calibration techniques can also mitigate the potential fairness issues within LLMs [551,552].”
(2024) · doi:10.1109/access.2024.3365742 · Cited by 934
Explainability for Large Language Models: A Survey
Zhao, H., Chen, H., Yang, F., Liu, N., Deng, H., Cai, H., Wang, S., Yin, D., & Du, M. (2024) · ACM Transactions on Intelligent Systems and Technology · doi:10.1145/3639372 · Cited by 681
“Post-hoc explainability methods can be used to analyze model feature importance patterns to identify biases or limitations in its behavior (Du et al., 2023).”
BackgroundMethodSelf-citationHarnessing the power of llms in practice: a survey on chatGPT and beyond
Yang, J., Jin, H., Tang, R., Han, X., Feng, Q., Jiang, H., Zhong, S., Yin, B., & Hu, X. (2024) · ACM Transactions on Knowledge Discovery from Data · doi:10.1145/3649506 · Cited by 644
Self-citationA Survey on Large Language Models: Applications, Challenges, Limitations, and Practical Usage
Hadi, M. U., tashi, Q. A., Qureshi, R., Shah, A., muneer, A., Irfan, M., Zafar, A., Shaikh, M. B., Akhtar, N., Wu, J., et al. (2023) · doi:10.36227/techrxiv.23589741.v1 · Cited by 370
Metadata
- Authors
- Du, M., He, F., Zou, N., Tao, D., & Hu, X.
- Affiliations
- New Jersey Institute of Technology (US)
- University of Edinburgh (GB)
- Texas A&M University (US)
- The University of Sydney (AU)
- Rice University (US)
- Venue
- Communications of the ACM
- Published
- 2023-12-21
- Volume
- 67
- Issue
- 1
- Pages
- 110-120
- ISSN
- 0001-0782 (ISSN-L)
- 1557-7317
- OpenAlex
- W4390043316
- Cited by
- 94
- References
- 40
- Citations by year
- 2026: 22 · 2025: 34 · 2024: 22 · 2023: 16
- FWCI
- 8.86
- Type
- journal article
- Language
- en
- Open access
- bronze
- doi.org
- Locations
- Communications of the ACM (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University) (cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Keywords
- shortcut learning, large language models, natural language understanding, robustness, model interpretability
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.