mlsanityData to 7 Oct 2026

Robust Pronoun Fidelity with English LLMs: Are they Reasoning, Repeating, or Just Biased?

arXiv 2404.03134 ↗First posted 4 Apr 2024cs.CL cs.CYcited by 17

On the map: LLM reasoning → Reasoning benchmarks and evaluation

Abstract (arXiv)

Robust, faithful and harm-free pronoun use for individuals is an important goal for language model development as their use increases, but prior work tends to study only one or two of these characteristics at a time. To measure progress towards the combined goal, we introduce the task of pronoun fidelity: given a context introducing a co-referring entity and pronoun, the task is to reuse the correct pronoun later. We present RUFF, a carefully-designed dataset of over 5 million instances to measure robust pronoun fidelity in English, and we evaluate 37 model variants from nine popular families, across architectures (encoder-only, decoder-only and encoder-decoder) and scales (11M-70B parameters). When an individual is introduced with a pronoun, models can mostly faithfully reuse this pronoun in the next sentence, but they are significantly worse with she/her/her, singular they and neopronouns. Moreover, models are easily distracted by non-adversarial sentences discussing other people; even one sentence with a distractor pronoun causes accuracy to drop on average by 34 percentage points. Our results show that pronoun fidelity is not robust, in a simple, naturalistic setting where humans achieve nearly 100% accuracy. We encourage researchers to bridge the gaps we find and to carefully evaluate reasoning in settings where superficial repetition might inflate perceptions of model performance.

Key facts. Cited by 17 papers in the map (data through 7 Oct 2026). Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning. Builds on 55 of its 81 references that are in the map. First posted on arXiv 4 Apr 2024.

Builds on

55 of 81 references.

Closest ancestors · 1

Same direction, newest first

  1. 2024Same Task, More Tokens: the Impact of Input Length on the Reasoning Performance of Large Language Models19 Feb

Other references · 54

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  2. 2019RoBERTa: A Robustly Optimized BERT Pretraining Approachcited by 13,266
  3. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 12,865
  4. 2022Large Language Models are Zero-Shot Reasonerscited by 4,323
  5. 2022OPT: Open Pre-trained Transformer Language Modelscited by 3,464
  6. 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,383
  7. 2022Scaling Instruction-Finetuned Language Modelscited by 2,997
  8. 2019ALBERT: A Lite BERT for Self-supervised Learning of Language Representationscited by 2,772
  9. 2023Pythia: A Suite for Analyzing Large Language Models Across Training and Scalingcited by 1,734
  10. 2016Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddingscited by 1,304
  11. 2022Large Language Models Are Human-Level Prompt Engineerscited by 1,020
  12. 2020Language (Technology) is Power: A Critical Survey of "Bias" in NLPcited by 951
38 more
  1. 2016Semantics derived automatically from language corpora contain human-like biasescited by 894
  2. 2023Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formattingcited by 778
  3. 2023The Flan Collection: Designing Data and Methods for Effective Instruction Tuningcited by 734
  4. 2018Gender Bias in Coreference Resolution: Evaluation and Debiasing Methodscited by 657
  5. 2023Are Emergent Abilities of Large Language Models a Mirage?cited by 515
  6. 2018Gender Bias in Coreference Resolutioncited by 438
  7. 2016A Corpus and Evaluation Framework for Deeper Understanding of Commonsense Storiescited by 413
  8. 2019On Measuring Social Biases in Sentence Encoderscited by 397
  9. 2019Measuring Bias in Contextualized Word Representationscited by 275
  10. 2019Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Themcited by 268
  11. 2019Gender Bias in Contextualized Word Embeddingscited by 203
  12. 2020Intrinsic Bias Metrics Do Not Correlate with Application Biascited by 148
  13. 2023Stop Uploading Test Data in Plain Text: Practical Strategies for Mitigating Data Contamination by Evaluation Benchmarkscited by 139
  14. 2021Harms of Gender Exclusivity and Challenges in Non-Binary Representation in Language Technologiescited by 128
  15. 2019Strategies for Structuring Story Generationcited by 124
  16. 2022Scaling Laws vs Model Architectures: How does Inductive Bias Influence Scaling?cited by 103
  17. 2018Context-Aware Neural Machine Translation Learns Anaphora Resolutioncited by 97
  18. 2019Toward Gender-Inclusive Coreference Resolutioncited by 84
  19. 2021Collecting a Large-Scale Gender Bias Dataset for Coreference Resolution and Machine Translationcited by 62
  20. 2022minicons: Enabling Flexible Behavioral and Representational Analyses of Transformer Language Modelscited by 58
  21. 2021Stereotype and Skew: Quantifying Gender Bias in Pre-trained and Fine-tuned Language Modelscited by 57
  22. 2023"I'm fully who I am": Towards Centering Transgender and Non-Binary Voices to Measure Biases in Open Language Generationcited by 56
  23. 2024Auxiliary task demands mask the capabilities of smaller language modelscited by 54
  24. 2022Welcome to the Modern World of Pronouns: Identity-Inclusive Natural Language Processing beyond Gendercited by 47
  25. 2023Prompting is not a substitute for probability measurements in large language modelscited by 43
  26. 2023MISGENDERED: Limits of Large Language Models in Understanding Pronounscited by 43
  27. 2023WinoQueer: A Community-in-the-Loop Benchmark for Anti-LGBTQ+ Bias in Large Language Modelscited by 42
  28. 2022Quantifying Social Biases Using Templates is Unreliablecited by 38
  29. 2021When Does Translation Require Context? A Data-driven, Multilingual Explorationcited by 36
  30. 2022Fewer Errors, but More Stereotypes? The Effect of Model Size on Gender Biascited by 33
  31. 2023MosaicBERT: A Bidirectional Encoder Optimized for Fast Pretrainingcited by 33
  32. 2021Back to Square One: Artifact Detection, Training and Commonsense Disentanglement in the Winograd Schemacited by 32
  33. 2023A Better Way to Do Masked Language Model Scoringcited by 27
  34. 2022Language model acceptability judgements are not always robust to contextcited by 22
  35. 2020The Sensitivity of Language Models and Humans to Winograd Schema Perturbationscited by 21
  36. 2017Dynamic Entity Representations in Neural Language Modelscited by 19
  37. 2022Using Natural Sentences for Understanding Biases in Language Modelscited by 19
  38. 2023What about em? How Commercial Machine Translation Fails to Handle (Neo-)Pronounscited by 19

50 most cited of 54 All 54 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2024: 4; 2025: 8; 2026: 5.

17 papers.

Sort:

By area:

  1. 2024Holmes: A Benchmark to Assess the Linguistic Competence of Language Modelscited by 26
  2. 2024WinoPron: Revisiting English Winogender Schemas for Consistency, Coverage, and Grammatical Casecited by 9
  3. 2026Whose Facts Win? LLM Source Preferences under Knowledge Conflictscited by 8
  4. 2025Colombian Waitresses y Jueces canadienses: Gender and Country Biases in Occupation Recommendations from LLMscited by 7
  5. 2025Agree to Disagree? A Meta-Evaluation of LLM Misgenderingcited by 6
  6. 2024The Lou Dataset -- Exploring the Impact of Gender-Fair Language in German Text Classificationcited by 5
  7. 2024NYT-Connections: A Deceptively Simple Text Classification Task that Stumps System-1 Thinkerscited by 4
  8. 2026Failure of contextual invariance in large language modelscited by 3
  9. 2025Uncertainty Quantification for Evaluating Machine Translation Biascited by 2
  10. 2025Adapting Psycholinguistic Research for LLMs: Gender-inclusive Language in a Coreference Contextcited by 1
  11. 2025KG-Augmented Executable CoT for Mathematical Codingcited by 1
  12. 2025SLAyiNG: A Diverse and Community-validated Dataset of Queer Slangcited by 1
5 more
  1. 2025Decoding Partial Differential Equations: Cross-Modal Adaptation of Decoder-only Models to PDEscited by 0
  2. 2025Teaching and Critiquing Conceptualization and Operationalization in NLPcited by 0
  3. 2026GRUFF: LLM Pronoun Fidelity, Reasoning, and Biases in Germancited by 0
  4. 2026A Mechanistic Understanding of Pronoun Fidelity in LLMscited by 0
  5. 2026Distinct dynamics of conceptual and referential disruptions in human reading and large language model processingcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Rethinking Faithfulness in LLMs: A Pairwise Context-Sensitive Perspective
  2. 2023Evaluating Concurrent Robustness of Language Models Across Diverse Challenge Sets
  3. 2025Conversational Speech Reveals Structural Robustness Failures in SpeechLLM Backbones
  4. 2026An Analysis of Training-Free Self-Reported Confidence in Language Models
  5. 2025FLUKE: A Linguistically-Driven and Task-Agnostic Framework for Robustness Evaluation
  6. 2022Underspecification in Language Modeling Tasks: A Causality-Informed Study of Gendered Pronoun Resolution
  7. 2023Trustworthy LLMs: a Survey and Guideline for Evaluating Large Language Models' Alignment
  8. 2025Calibrating Verbalized Confidence with Self-Generated Distractors