mlsanityData to 7 Oct 2026

What Will it Take to Fix Benchmarking in Natural Language Understanding?

arXiv 2104.02145 ↗First posted 5 Apr 2021cs.CLcited by 200

Not on the map. Why?

Abstract (arXiv)

Evaluation for many natural language understanding (NLU) tasks is broken: Unreliable and biased systems score so highly on standard benchmarks that there is little room for researchers who develop better systems to demonstrate their improvements. The recent trend to abandon IID benchmarks in favor of adversarially-constructed, out-of-distribution test sets ensures that current models will perform poorly, but ultimately only obscures the abilities that we want our benchmarks to measure. In this position paper, we lay out four criteria that we argue NLU benchmarks should meet. We argue most current benchmarks fail at these criteria, and that adversarial data collection does not meaningfully address the causes of these failures. Instead, restoring a healthy evaluation ecosystem will require significant progress in the design of benchmark datasets, the reliability with which they are annotated, their size, and the ways they handle social bias.

Key facts. Cited by 200 papers in the map (data through 7 Oct 2026). Builds on 45 of its 75 references that are in the map. First posted on arXiv 5 Apr 2021.

Builds on

45 of 75 references.

References · 45

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  2. 2020Language Models are Few-Shot Learnerscited by 32,199
  3. 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,491
  4. 2019RoBERTa: A Robustly Optimized BERT Pretraining Approachcited by 13,266
  5. 2018GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understandingcited by 4,854
  6. 2016SQuAD: 100,000+ Questions for Machine Comprehension of Textcited by 4,115
  7. 2020DeBERTa: Decoding-enhanced BERT with Disentangled Attentioncited by 2,157
  8. 2015A large annotated corpus for learning natural language inferencecited by 1,837
  9. 2019SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systemscited by 1,718
  10. 2018Know What You Don't Know: Unanswerable Questions for SQuADcited by 1,538
  11. 2018Datasheets for Datasetscited by 1,371
  12. 2016Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddingscited by 1,304
33 more
  1. 2020Beyond Accuracy: Behavioral Testing of NLP models with CheckListcited by 979
  2. 2016The LAMBADA dataset: Word prediction requiring a broad discourse contextcited by 960
  3. 2020Language (Technology) is Power: A Critical Survey of "Bias" in NLPcited by 951
  4. 2019Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inferencecited by 864
  5. 2019Adversarial NLI: A New Benchmark for Natural Language Understandingcited by 798
  6. 2020It's Not Just Size That Matters: Small Language Models Are Also Few-Shot Learnerscited by 668
  7. 2017Adversarial Examples for Evaluating Reading Comprehension Systemscited by 568
  8. 2018Gender Bias in Coreference Resolutioncited by 438
  9. 2018SWAG: A Large-Scale Adversarial Dataset for Grounded Commonsense Inferencecited by 333
  10. 2019Cosmos QA: Machine Reading Comprehension with Contextual Commonsense Reasoningcited by 281
  11. 2019Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Themcited by 268
  12. 2020Experience Grounds Languagecited by 241
  13. 2018Examining Gender and Race Bias in Two Hundred Sentiment Analysis Systemscited by 234
  14. 2019Probing Neural Network Comprehension of Natural Language Argumentscited by 212
  15. 2018Stress Test Evaluation for Natural Language Inferencecited by 211
  16. 2019Show Your Work: Improved Reporting of Experimental Resultscited by 164
  17. 2020Adversarial Filters of Dataset Biasescited by 159
  18. 2020With Little Power Comes Great Responsibilitycited by 138
  19. 2020Towards Accountability for Machine Learning Datasets: Practices from Software Engineering and Infrastructurecited by 138
  20. 2020Intermediate-Task Transfer Learning with Pretrained Models for Natural Language Understanding: When and Why Does It Work?cited by 127
  21. 2020UnQovering Stereotyping Biases via Underspecified Questionscited by 124
  22. 2020OCNLI: Original Chinese Natural Language Inferencecited by 86
  23. 2018Performance Impact Caused by Hidden Bias of Training Data for Recognizing Textual Entailmentcited by 58
  24. 2019Assessing the Benchmarking Capacity of Machine Reading Comprehension Datasetscited by 46
  25. 2019Human vs. Muppet: A Conservative Estimate of Human Performance on the GLUE Benchmarkcited by 32
  26. 2019What does BERT Learn from Multiple-Choice Reading Comprehension Datasets?cited by 30
  27. 2020Utility is in the Eye of the User: A Critique of NLP Leaderboardscited by 28
  28. 2017Towards Linguistically Generalizable NLP Systems: A Workshop and Shared Taskcited by 24
  29. 2019What Question Answering can Learn from Trivia Nerdscited by 21
  30. 2019Metrology for AI: From Benchmarks to Instrumentscited by 12
  31. 2020Asking Crowdworkers to Write Entailment Examples: The Best of Bad Optionscited by 8
  32. 2019Large-scale Pretraining for Neural Machine Translation with Tens of Billions of Sentence Pairscited by 4
  33. 2020TuringAdvice: A Generative and Dynamic Evaluation of Language Usecited by 4

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2021: 21; 2022: 33; 2023: 27; 2024: 23; 2025: 32; 2026: 64.

200 papers.

Sort:

By area:

  1. 2023SWE-bench: Can Language Models Resolve Real-World GitHub Issues?cited by 3,193
  2. 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,748
  3. 2023PaLM 2 Technical Reportcited by 1,154
  4. 2021AI and the Everything in the Whole Wide World Benchmarkcited by 340
  5. 2021LexGLUE: A Benchmark Dataset for Legal Language Understanding in Englishcited by 302
  6. 2021Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Modelscited by 227
  7. 2021QuALITY: Question Answering with Long Input Texts, Yes!cited by 204
  8. 2022WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creationcited by 200
  9. 2021Masked Language Modeling and the Distributional Hypothesis: Order Word Matters Pre-training for Littlecited by 176
  10. 2022Repairing the Cracked Foundation: A Survey of Obstacles in Evaluation Practices for Generated Textcited by 154
  11. 2021Scaling Up Influence Functionscited by 130
  12. 2021KLUE: Korean Language Understanding Evaluationcited by 113
38 more
  1. 2023Generative Language Models Exhibit Social Identity Biasescited by 107
  2. 2025Measuring what Matters: Construct Validity in Large Language Model Benchmarkscited by 102
  3. 2025TRAIL: Trace Reasoning and Agentic Issue Localizationcited by 98
  4. 2023Revisiting Out-of-distribution Robustness in NLP: Benchmark, Analysis, and LLMs Evaluationscited by 96
  5. 2021CREAK: A Dataset for Commonsense Reasoning over Entity Knowledgecited by 88
  6. 2023Anchor Points: Benchmarking Models with Much Fewer Examplescited by 84
  7. 2022Mapping global dynamics of benchmark creation and saturation in artificial intelligencecited by 71
  8. 2023The Song Describer Dataset: a Corpus of Audio Captions for Music-and-Language Evaluationcited by 68
  9. 2023Do LLMs Understand Social Knowledge? Evaluating the Sociability of Large Language Models with SocKET Benchmarkcited by 62
  10. 2025Can We Trust AI Benchmarks? An Interdisciplinary Review of Current Issues in AI Evaluationcited by 57
  11. 2024Auxiliary task demands mask the capabilities of smaller language modelscited by 54
  12. 2022FairLex: A Multilingual Benchmark for Evaluating Fairness in Legal Text Processingcited by 48
  13. 2022Evaluation Gaps in Machine Learning Practicecited by 48
  14. 2025The SWE-Bench Illusion: When State-of-the-Art LLMs Remember Instead of Reasoncited by 47
  15. 2024VariErr NLI: Separating Annotation Error from Human Label Variationcited by 42
  16. 2025Toward an Evaluation Science for Generative AI Systemscited by 42
  17. 2023Language acquisition: do children and language models follow similar learning stages?cited by 40
  18. 2021The Dangers of Underclaiming: Reasons for Caution When Reporting How NLP Systems Failcited by 38
  19. 2025Answer Matching Outperforms Multiple Choice for Language Model Evaluationcited by 37
  20. 2022GEMv2: Multilingual NLG Benchmarking in a Single Line of Codecited by 36
  21. 2022Validating Large Language Models with ReLMcited by 33
  22. 2021Models in the Loop: Aiding Crowdworkers with Generative Annotation Assistantscited by 31
  23. 2021Retrieval-guided Counterfactual Generation for QAcited by 28
  24. 2022What Do NLP Researchers Believe? Results of the NLP Community Metasurveycited by 28
  25. 2025RefactorBench: Evaluating Stateful Reasoning in Language Agents Through Codecited by 28
  26. 2024ECBD: Evidence-Centered Benchmark Design for NLPcited by 27
  27. 2024Inherent Trade-Offs between Diversity and Stability in Multi-Task Benchmarkscited by 26
  28. 2023It Takes Two to Tango: Navigating Conceptualizations of NLP Tasks and Measurements of Performancecited by 25
  29. 2024Automated Evaluation of Retrieval-Augmented Language Models with Task-Specific Exam Generationcited by 25
  30. 2025Multi-turn Evaluation of Anthropomorphic Behaviours in Large Language Modelscited by 25
  31. 2021Analyzing Dynamic Adversarial Training Data in the Limitcited by 24
  32. 2022Law Informs Code: A Legal Informatics Approach to Aligning Artificial Intelligence with Humanscited by 23
  33. 2024RuBLiMP: Russian Benchmark of Linguistic Minimal Pairscited by 23
  34. 2024The Vulnerability of Language Model Benchmarks: Do They Accurately Reflect True LLM Performance?cited by 22
  35. 2022Evaluate & Evaluation on the Hub: Better Best Practices for Data and Model Measurementscited by 21
  36. 2024Real Risks of Fake Data: Synthetic Data, Diversity-Washing and Consent Circumventioncited by 21
  37. 2026Brittlebench: Quantifying LLM robustness via prompt sensitivitycited by 19
  38. 2021Automatic Construction of Evaluation Suites for Natural Language Generation Datasetscited by 18

50 most cited of 200 All 200 (JSON)