The DOI matches its Crossref record; 20 of the 20 references it cites were found in a registry or an index.
HellaSwag: Can a Machine Really Finish Your Sentence?
Open accessAbstract
Recent work by Zellers et al. (2018) introduced a new task of commonsense natural language inference: given an event description such as "A woman sits at a piano," a machine must select the most likely followup: "She sets her fingers on the keys."With the introduction of BERT (Devlin et al., 2018), near human-level performance was reached.Does this mean that machines can perform human level commonsense inference?In this paper, we show that commonsense inference still proves di cult for even stateof-the-art models, by presenting HellaSwag, a new challenge dataset.Though its questions are trivial for humans (°95% accuracy), state-of-the-art models struggle ( †48%).We achieve this via Adversarial Filtering (AF), a data collection paradigm wherein a series of discriminators iteratively select an adversarial set of machine-generated wrong answers.AF proves to be surprisingly robust.The key insight is to scale up the length and complexity of the dataset examples towards a critical 'Goldilocks' zone wherein generated text is ridiculous to humans, yet often misclassified by state-of-the-art models.Our construction of HellaSwag, and its resulting di culty, sheds light on the inner workings of deep pretrained models.More broadly, it suggests a new path forward for NLP research, in which benchmarks co-evolve with the evolving state-of-the-art in an adversarial way, so as to present ever-harder challenges.
Overview
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
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?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“There is no 'ground-truth' implementation from the original benchmark authors, and many popular benchmarks (Paperno et al., 2016; Clark et al., 2018; Zellers et al., 2019) that have been 'retrofitted' for compatibility with autoregressive LMs.”
From Lessons from the Trenches on Reproducible Evaluation of Language Models (2024) · cites this paper as Zellers et al., 2019
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Second, benchmarks are often adversarially selected for tasks that current models struggle with compared to humans, 1 biasing the comparison to human performance.”
From Measuring AI Ability to Complete Long Software Tasks (2025) · cites this paper as [4]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“During RLHF fine-tuning, we observe performance regressions compared to GPT-3 on certain public NLP datasets, notably SQuAD (Rajpurkar et al., 2018), DROP (Dua et al., 2019), HellaSwag (Zellers et al., 2019), and WMT 2015 French to English translation (Bojar et al., 2015).”
From Training language models to follow instructions with human feedback (2022) · cites this paper as Zellers et al., 2019
The cited source says what the claim attributes to it.
In the citing paper
“Traditional language model benchmarks, such as HellaSwag (Zellers et al., 2019), WinoGrande (Sakaguchi et al., 2021), MMLU (Hendrycks et al., 2021), GSM-8K (Cobbe et al., 2021), and BIG-Bench (Srivastava et al., 2023), primarily assess LLMs on tasks like commonsense reasoning and standardized exams.”
From LLM Self-Correction with DeCRIM: Decompose, Critique, and Refine for Enhanced Following of Instructions with Multiple Constraints (2024) · cites this paper as Zellers et al., 2019
The cited source says what the claim attributes to it.
In the citing paper
“for evaluating the following tasks that measure common sense reasoning: LAMBADA (OpenAI version in LM Evaluation Harness) (Paperno et al., 2016), HellaSwag (Zellers et al., 2019), PIQA (Bisk et al., 2020), ARC-challenge, ARC-easy (Clark et al., 2018), WinoGrande (Sakaguchi et al., 2021).”
From xLSTM: Extended Long Short-Term Memory (2024) · cites this paper as Zellers et al., 2019
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
Deep Contextualized Word Representations
Peters, M., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., & Zettlemoyer, L. (2018) · Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) · doi:10.18653/v1/n18-1202 · Cited by 11,983
Williams, S., Waterman, A., & Patterson, D. A. (2009) · Communications of the ACM · doi:10.1145/1498765.1498785 · Cited by 2,798
Zhu, Y., Kiros, R., Zemel, R., Salakhutdinov, R., Urtasun, R., Torralba, A., & Fidler, S. (2015) · 2015 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2015.11 · Cited by 2,264
Adversarial Examples for Evaluating Reading Comprehension Systems
Jia, R., & Liang, P. (2017) · Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d17-1215 · Cited by 1,410
Enhanced LSTM for Natural Language Inference
Chen, Q., Zhu, X., Ling, Z.-H., Wei, S., Jiang, H., & Inkpen, D. Z. (2017) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/p17-1152 · Cited by 1,315
The Curious Case of Neural Text Degeneration
Holtzman, A., Buys, J., Du, L., Forbes, M., & Choi, Y. (2019) · arXiv · doi:10.48550/arxiv.1904.09751 · Cited by 1,304
Self-citationDense-Captioning Events in Videos
Krishna, R., Hata, K., Ren, F., Fei-Fei, L., & Niebles, J. C. (2017) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.83 · Cited by 1,279
Annotation Artifacts in Natural Language Inference Data
Gururangan, S., Swayamdipta, S., Levy, O., Schwartz, R., Bowman, S., & Smith, N. A. (2018) · Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers) · doi:10.18653/v1/n18-2017 · Cited by 907
Research works printed but not linked: 2
- Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2017. Bag of tricks for e cient text classification. In Proceedings of the 15th Confer- ence of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, volume 2, pages 427-431.
- Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word repre- sentations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), volume 1, pages 2227-2237.
Cited by
Sorted by citations
Items 901–908 of 908
A2A—A Mechanism Driven Framework for Interaction Based Agent Evaluation
Zhang, H., Zhang, R., Li, Y., Zhao, K., Song, Y., Song, Z., & Liu, G. (2026) · International Conference on Artificial Intelligence and Computing Power · doi:10.1145/3843915.3843923
“Traditional approaches rely on static benchmarks such as MMLU [4] and HellaSwag [12], which focus on single-turn outcome correctness.”
WSM: Decay-Free Learning Rate Schedule via Checkpoint Merging for LLM Pre-training
Tian, C., Wang, J., Zhao, Q., Chen, K., Liu, J., Liu, Z., Mao, J., Zhao, W. X., Zhang, Z., & Zhou, J. (2025) · arXiv · doi:10.48550/arxiv.2507.17634
“…et al., 2024), OpenBookQA (Mihaylov et al., 2018), BBH (Suzgun et al., 2023), WorldSense (Hong et al., 2025), PIQA (Bisk et al., 2020), hellaswag (Zellers et al., 2019) and KOR-Bench (Ma et al., 2025)) (b) Language Understanding (e.g., race (Lai et al., 2017), SQuAD 2.0 (Rajpurkar et al., 2018),…”
InfluentialEfficient Benchmarking (of Language Models)
Perlitz, Y., Bandel, E., Gera, A., Arviv, O., Ein-Dor, L., Shnarch, E., Slonim, N., Shmueli-Scheuer, M., & Choshen, L. (2023) · arXiv · doi:10.48550/arxiv.2308.11696
LittleBit: Ultra Low-Bit Quantization via Latent Factorization
Lee, B., Kim, D., You, Y., & Kim, Y. (2025) · arXiv · doi:10.48550/arxiv.2506.13771 · Cited by 7
“We also assess zero-shot accuracy on common reasoning benchmarks: WinoGrande [57], OpenBookQA (OBQA) [58], HellaSwag [59], BoolQ [60], ARC-Easy (ARC-e), ARC-Challenge (ARC-c) [61], and PIQA [62].”
Defending Against Neural Fake News
Zellers, R., Holtzman, A., Rashkin, H., Bisk, Y., Farhadi, A., Roesner, F., & Choi, Y. (2019) · arXiv · doi:10.48550/arxiv.1905.12616 · Cited by 569
Self-citationUnderspecification Presents Challenges for Credibility in Modern Machine Learning
D'Amour, A., Heller, K., Moldovan, D., Adlam, B., Alipanahi, B., Beutel, A., Chen, C., Deaton, J., Eisenstein, J., Hoffman, M. D., et al. (2020) · arXiv · doi:10.48550/arxiv.2011.03395 · Cited by 398
NormFormer: Improved Transformer Pretraining with Extra Normalization
Shleifer, S., Weston, J., & Ott, M. (2021) · arXiv · doi:10.48550/arxiv.2110.09456 · Cited by 24
Byte Latent Transformer: Patches Scale Better Than Tokens
Pagnoni, A., Pasunuru, R., Rodriguez, P., Nguyen, J., Muller, B., Li, M., Zhou, C., Yu, L., Weston, J., Zettlemoyer, L., et al. (2024) · arXiv · doi:10.48550/arxiv.2412.09871
Self-citation
Metadata
- Authors
- Zellers, R., Holtzman, A., Bisk, Y., Farhadi, A., & Choi, Y.
- Affiliations
- University of Washington (US)
- Venue
- Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics
- Published
- 2019
- Pages
- 4791-4800
- ISSN
- 0736-587X (ISSN-L)
- OpenAlex
- W2946609015
- Cited by
- 830
- References
- 20
- Citations by year
- 2026: 200 · 2025: 263 · 2024: 109 · 2023: 93 · 2022: 50 · 2021: 61 · 2020: 44 · 2019: 10
- FWCI
- 15.78
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- Locations
- Proceedings of the conference - Association for Computational Linguistics. Meeting (publishedVersion, cc-by)
- Funders
- National Science Foundation (1703166, 1256082, 1637479, 1524371)
- Nvidia
- Allen Institute for Artificial Intelligence
- Defense Advanced Research Projects Agency
- Intelligence Advanced Research Projects Activity
- Army Research Office (W911NF-15-1-, W911NF)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
- Keywords
- commonsense reasoning, natural language inference, AI-generated text, BERT, benchmark design, pre-trained language models
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
