PaperlayerView exact revision on arXiv

Source-linked AI summary

Investigating the Impact of Learner Errors on Social Bias in Large Language Models

Federico Marcuzzi, Martin Christoph, Iryna Gurevych, Alla Rozovskays

arXiv:2610.29896v1cs.CL

TL;DR

Existing bias benchmarks mostly use well-formed English, while L2 writing more often contains learner errors. This paper injects controlled errors into bias benchmarks and finds that errors generally increase bias as their rate rises, with preposition errors especially affecting stereotype alignment.

  • Problem

    Existing bias benchmarks mostly use well-formed English, although L2 speakers outnumber L1 speakers globally and L2 writing more often contains learner errors.

    Source: §1 · p. 1

  • Method

    The paper injects controlled learner errors into bias benchmarks, varying error type, frequency, and prompt location.

    Source: §1 · p. 2

  • Results

    Across benchmarks, errors generally increased stereotype alignment and demographic preference with error rate; preposition errors most affected stereotype alignment, and error-aware fine-tuning mitigated increases.

    Source: §4.2 · p. 6, §5 · p. 8

  • Takeaways & Limitations

    Evaluations using only well-formed prompts may underestimate bias experienced by linguistically diverse users, and improving robustness to variation can reduce that bias.

    Source: §5 · pp. 8–9

    Rule-based synthetic errors calibrated to learner corpora may not capture the complexity of natural L2 writing.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Existing social bias benchmarks evaluate large language models mostly on fluent, well-formed English, reflecting native speaker usage and leaving non-native writers unrepresented. To address this gap, we investigate how grammatical errors by non-native writers affect social bias. We develop a controlled framework that injects common learner errors into prompts and apply it to three bias benchmarks covering stereotype alignment and demographic preference across gender, race, and religion. We find that grammatical errors increase bias, and the effect grows with the error rate. Among the four common error types we consider, preposition errors produce the largest and most consistent increase in stereotype alignment. Our analysis suggests that errors increase bias when they are more likely to interfere with comprehension. Finally, fine-tuning on erroneous texts reduces the bias increase caused by grammatical errors. Our findings show that bias evaluations based exclusively on well-formed prompts may underestimate the bias experienced by users whose English contains grammatical errors.

1 Introduction

Existing bias benchmarks mostly use well-formed English, leaving the grammatical errors common in L2 writing underexamined. The paper introduces controlled prompt perturbations to test how these errors affect social bias in LLMs.

1 Introduction

  • Existing bias work mostly measures models with well-formed text, implicitly assuming native-speaker users despite L2 speakers outnumbering L1 speakers globally.L2 writing is also more likely to contain language-learner grammatical errors.Source: §1 · p. 1
  • Prior work links prompt noise to altered LLM behavior, motivating investigation of whether grammatical errors interact with existing biases.The passage presents altered token probabilities, reduced clarity, and increased ambiguity as possible mechanisms.Source: §1 · p. 1
  • The framework injects common grammatical errors into existing benchmarks while controlling their type, frequency, and prompt location.The benchmarks cover stereotype alignment and demographic preference across tasks and gender, race, and religion.Source: §1 · pp. 1–2
  • The study asks whether L2 prompt errors increase bias, how error type and location affect it, and whether error-aware fine-tuning can mitigate the effect.Source: §1 · p. 2

2 Related Work

Prior research measures social bias through curated tasks, but standard benchmarks use well-formed English and prior linguistic-variation studies do not measure benchmark-elicited demographic bias. Synthetic-error research supplies methods for simulating learner writing, including controlled linguistic rules.

2 Related Work

  • Social bias is measured with curated benchmarks targeting particular representational or allocational harms through tasks such as association tests, coreference, question answering, and decisions.Source: §2 · p. 2
  • Existing bias benchmarks use standard well-formed English, and prior linguistic-variation work does not measure demographic bias elicited by those benchmarks.Prior studies instead examine stereotyping, discrimination, classification accuracy, or false detection of L2 essays as machine-generated.Source: §2 · p. 2
  • Prompt-noise research finds performance drops, but targets task accuracy rather than social bias (Wang et al., 2024).Source: §2 · p. 2
  • Synthetic learner-error methods include noise-based, translation-based, and linguistically inspired approaches, with realism often grounded in learner corpora (Náplava and Straka, 2019).The paper adopts linguistic rules and part-of-speech information to control error type, frequency, and location.Source: §2 · pp. 2–3

3 Methodology

The methodology perturbs established bias benchmarks rather than designing new prompts, enabling controlled comparisons with clean prompts. It covers multiple bias forms and tasks, and uses rule-based synthetic errors with controlled types, locations, and rates.

3 Methodology

  • Controlled perturbations of established, error-free benchmarks isolate linguistic noise as the independent variable and enable comparison with baseline bias.Synthetic inputs let the authors control error type, rate, and location, and all models receive identical pre-generated prompts.Source: §3 · p. 3

3.1 Benchmarks

  • The benchmarks cover stereotype alignment and demographic preference across coreference, question answering, and decision-making in gender, race, and religion.Source: §3.1 · p. 3
  • WinoBias measures gender stereotype alignment in coreference using pro- and anti-stereotypical pairs; its historical bias score is the accuracy difference between them.Source: §3.1 · p. 3
  • BBQ measures stereotype alignment in gender, race, and religion; its ambiguous-context score compares pro- with anti-stereotypical answers after neutral options are removed.Source: §3.1 · p. 3
  • DiscrimEvalGen measures demographic preference in decisions by comparing historically advantaged and disadvantaged groups using a demographic parity gap.Source: §3.1 · p. 3

3.2 Generating Synthetic Errors

  • Rule-based perturbations simulate determiner, preposition, verb, and spelling errors, with eligible tokens and prompt locations controlled and protected content left unchanged.Operation and subtype distributions are drawn from learner corpora; spelling perturbations include close dictionary words and character-level misspellings.Source: §3.2 · p. 4
  • Error rate is token-level Levenshtein distance divided by unprotected-token count; tested levels are 5%, 10%, 20%, and 30%, calibrated to within one percentage point when reachable.The selected levels span rates observed in learner corpora, typically 10–20% of tokens.Source: §3.2 · p. 5

4 Experiments and Results

Across the experiments, grammatical errors generally increased bias, with effects varying by error type, location, and task. The analysis links the increase to impaired comprehension, while error-aware fine-tuning reduced it on the evaluated bias benchmarks.

4.1 Experimental Setup

  • The default setup combined four error types in the input and used gender, the demographic axis shared across all three benchmarks.The study evaluated five open-weight models and used both option orders with five generations per prompt.Source: §4.1 · p. 5

4.2 Key Results

  • Across benchmarks, grammatical errors increased stereotype alignment and demographic preference, with stronger effects at higher error levels; race and religion shifts were weaker and less consistent.On DE-Gen, errors reinforced a pre-existing preference for disadvantaged groups rather than shifting preference toward advantaged groups.Source: §4.2 · p. 6

4.3 The Effect of the Error Type

  • On WinoBias, preposition errors produced the largest stereotype-alignment increase for every model, exceeding the combined error types at the same rate.On BBQ, preposition and verb errors tended to have the strongest effects, while DE-Gen showed no consistent error-type trend; comprehension-degrading errors were associated with larger bias increases (r = 0.76).Source: §4.3 · p. 6, Figure 2 · p. 6

4.4 The Effect of the Error Location

  • On WinoBias, input errors amplified bias, whereas instruction-only errors usually did not; errors in both locations produced values close to, and often slightly below, input-only errors.The authors hypothesize that errors in both locations impair task comprehension and narrow the bias gap; location effects were less pronounced on BBQ and inconsistent on DE-Gen.Source: §4.4 · pp. 6–7, Figure 3 · p. 7

4.5 Why L2 Errors Exacerbate Social Bias

  • Under errors, every model lost accuracy on both WinoBias sentence types, but losses concentrated on anti-stereotypical sentences, consistent with fallback to stereotypical associations when comprehension degrades.The additional loss did not occur on tasks where no stereotype offered an alternative answer.Source: §4.5 · p. 7
  • In an example with identical errors on both WinoBias versions, the model retained the pro-stereotypical answer but resolved the anti-stereotypical pronoun using occupational stereotypes or role heuristics.The authors’ qualitative analysis interprets the stereotype as replacing sentence structure as the cue for pronoun resolution.Source: §4.5 · p. 7, Table 3 · p. 8

4.6 Mitigation via Error-Aware Fine-Tuning

  • Fine-tuning on errored prompts reduced error-induced bias growth on BBQ and lessened it on WinoBias, with better retained accuracy on anti-stereotypical sentences.The robustness did not transfer to MMLU; clean accuracy there remained unchanged.Source: §4.6 · p. 8

5 Conclusion

Grammatical errors generally increase social bias, with larger effects at higher error rates. The conclusion highlights preposition errors, comprehension-related effects, and error-aware fine-tuning as key findings.

5 Conclusion

  • Grammatical errors generally increase social bias, and the effect grows with the error rate.The increase is associated with a disproportionate loss of accuracy on anti-stereotypical examples, suggesting degraded comprehension leads models to rely more on stereotypes.Source: §5 · p. 8
  • Preposition errors have the largest effect on stereotype alignment among the four error types, which the authors attribute to semantic change.The increase in bias is associated with a disproportionate loss of accuracy on anti-stereotypical examples, suggesting degraded comprehension leads to greater reliance on stereotypical associations.Source: §5 · p. 8
  • Error-aware fine-tuning removes the bias increase on BBQ and reduces it on WinoBias.Source: §5 · p. 8
  • The findings identify linguistic variation as important in evaluating social bias and suggest robustness improvements can reduce bias encountered by linguistically diverse users.Source: §5 · pp. 8–9

Limitations

The study uses rule-based synthetic perturbations and standard template-based benchmarks, with a limited set of error categories and open-weight models.

Limitations

  • Rule-based synthetic perturbations are calibrated to learner-corpus distributions, but their realism was not validated with human judgments or L2-written text.Natural L2 writing can include interlanguage phenomena, discourse-level awkwardness, word-order permutations, and L1-specific syntactic transfers that the rules may not capture.Source: Limitations · p. 9
  • The analysis covers four error categories: determiners, prepositions, verbs, and spelling.The authors identify punctuation, lexical-choice errors, and larger multi-word phrasing anomalies as areas for future study.Source: Limitations · p. 9
  • The template-based benchmarks cover gender, race, and religion but not the full complexity of intersectional bias or open-ended chat.Source: Limitations · p. 9
  • The evaluation uses open-weight models up to 32B parameters, and error-aware fine-tuning is tested on a single model.The authors state that budget constraints motivated these choices.Source: Limitations · p. 9

Ethical Considerations

The authors caution that their results cover only four learner-error types and three benchmarks, and should not be treated as exhaustive evidence about real L2 writing or other tasks and demographic groups.

Ethical Considerations

  • The framework covers four learner-error types, while real L2 writing contains many more.A small measured bias increase under these errors does not rule out a larger increase in real use.Source: Ethical Considerations · p. 9
  • The findings rest on three benchmarks, so generalization to other tasks and demographic groups requires caution.Source: Ethical Considerations · p. 9
  • The authors frame the findings as a model shortcoming, not a reason for L2 speakers to adapt their writing.The released framework could also be used to craft prompts that elicit more stereotypical answers.Source: Ethical Considerations · p. 9

A Evaluation Benchmarks

The benchmarks cover stereotype alignment and demographic preference across coreference, question answering, and decision-making. The appendix also describes capability evaluations and a separate dataset used for fine-tuning.

A Evaluation Benchmarks

  • WinoBias measures gender stereotype alignment in coreference using two occupations and a gendered pronoun, with pro- and anti-stereotypical versions.The test set also supports resolution accuracy; error injection leaves occupations, pronouns, and answer options untouched.Source: A Evaluation Benchmarks · p. 13
  • BBQ measures stereotype selection on its ambiguous split; its disambiguated split is used only to measure accuracy under errors.The benchmark includes gender, race, and religion, and demographic mentions, questions, and answer options are protected from error injection.Source: A Evaluation Benchmarks · p. 13
  • DiscrimEvalGen measures demographic preference by comparing decisions between male and female, Black and White, or Muslim and Christian groups.There is no correct answer because the groups are interchangeable; positive bias values indicate preference for the historically disadvantaged group.Source: A Evaluation Benchmarks · p. 13
  • MMLU tests general knowledge and reasoning across 57 subjects, while WinoGrande is used only for error-aware fine-tuning.WinoGrande training uses 40,398 items and its 1,267-item validation split is a held-out check; items matching selected WinoBias content are removed.Source: A Evaluation Benchmarks · p. 13

B.1 Error Types and Protected Spans

The framework simulates four common learner-error types while protecting benchmark-critical content and counterfactual contrasts. Error rates are calibrated to target levels separately by prompt part.

B.1 Error Types and Protected Spans

  • Determiner, preposition, verb, and spelling errors are generated through type-specific edits to eligible tokens.Verb subtypes follow W&I+LOCNESS ERRANT frequencies; spelling candidates use Aspell suggestions, and modified tokens are frozen against further edits.Source: §B.1 · p. 14
  • Protected spans include demographic entities, WinoBias pronouns, questions, answer options, and the boxed answer format.Counterfactual prompt pairs receive identical edits, preserving their difference in demographic attribute; shared instructions are perturbed independently.Source: §B.1 · p. 14
  • The combined error-type proportions are based on annotated edits from FCE, NUCLE, and W&I+LOCNESS.Table B.2 summarizes the share of edits by error type and the corresponding ERRANT categories.Source: Table B.2 · p. 14

B.2 Calibration of the Error Rate

  • Target error rates are calibrated separately for each benchmark, error type, and prompt part until the mean rate is within one percentage point of its target.Unattainable levels are discarded; combined errors split the target among types using FCE shares, and errors in both prompt parts are the union of the separately calibrated conditions.Source: §B.2 · p. 14

C.1 Comprehension Loss by Error Type

Across error types, those associated with greater comprehension loss also show greater bias increases. The additional accuracy loss is concentrated on anti-stereotypical WinoBias items, where the correct answer contradicts the stereotype.

C.1 Comprehension Loss by Error Type

  • Preposition errors increase WinoBias bias by 8.2 points, versus 5.0 for all four error types combined, despite lower comprehension loss (4.1 vs. 4.8 points).Across 80 model, error-type, and level combinations, comprehension loss and bias increase correlate at r = 0.76 (p < 0.001); preposition errors disproportionately reduce anti-stereotypical accuracy (12.4 vs. 9.8 points).Source: Table C.1 · p. 15, §C.1 · p. 15

C.2 Accuracy on Tasks Not Tied to a Stereotype

  • At 30% error rate, accuracy losses are 2.9–4.6 points on MMLU and 0.0–4.6 points on disambiguated BBQ, comparable to 3.5–5.8 points on pro-stereotypical WinoBias items.Source: §C.2 · p. 15
  • The reported loss is roughly twice as large (8.7 to 11.3 points); errors degrade comprehension only moderately in general, with additional loss where a stereotype offers an alternative answer.Source: §C.2 · p. 15

C.3 Error-Aware Fine-Tuning

Error-aware fine-tuning reduced the bias increase from grammatical errors on WinoBias and BBQ, though the benefit varied by benchmark. The comparison isolates exposure to errors during training from other effects of fine-tuning.

C.3 Error-Aware Fine-Tuning

  • The training set began with 5,000 filtered WinoGrande items rendered in both option orders, yielding 8,978 examples with correct model-generated reasoning targets.For FT errors, the clean prompts were replaced with errored versions; each item's prompts shared edits and an error level sampled from 5%, 10%, 20%, or 30%.Source: §C.3 · p. 15
  • Both models were fine-tuned with LoRA for one epoch, using a learning rate of 10−4 and an effective batch size of 16.The loss was computed on completions only.Source: §C.3 · p. 15
  • On WinoBias and BBQ, FT errors was significantly less biased at ˜ε = 30% than FT clean and Base, and its bias increase was significantly smaller than FT clean.Against Base, the increase was significantly smaller only on BBQ; on DE-Gen, neither the bias at ˜ε = 30% nor the increase differed significantly.Source: Table C.3 · p. 16, §C.3 · p. 16
  • FT clean and FT errors shared targets, prompt order, hyperparameters, and random seed; only FT errors used prompts containing grammatical errors.Both models were evaluated on the same prompts and decoding settings as the base model.Source: §C.3 · p. 16

D Reproducibility

The reproducibility materials specify inference settings and statistical testing, report computational and software details, and disclose AI-assistant use. Appendix E supplies additional tables and prompt examples.

D.1 Settings and Hyperparameters

  • Inference used five generations per prompt, with temperature 0.7, top-p 0.9, a 2,000-token limit, and fixed seeds for sampling and error injection.Prompts were plain text without a chat template; answers were determined by the most frequently boxed option.Source: §D.1 · p. 16

D.2 Statistical Significance

  • Differences were assessed with a paired approximate randomization test using 10,000 permutations and a two-sided α = 0.05 threshold.For model comparisons of error effects, each item's value was its change from the clean to the errored prompt.Source: §D.2 · p. 16

D.3 Computational Cost and Infrastructure

  • Experiments used NVIDIA H200 GPUs, with each model served on one GPU, and totaled approximately 700 GPU-hours including preliminary runs.Inference used vLLM.Source: §D.3 · pp. 16–17

D.4 Software and Licenses

  • The authors list the software versions and licenses used for inference, fine-tuning, and error generation, and release their code under Apache-2.0.The software list includes vLLM, Transformers, PEFT, spaCy, and GNU Aspell.Source: §D.4 · p. 17

D.5 Use of AI Assistants

  • AI assistants were used to debug and refactor code and proofread the manuscript; the authors validated the refactored code against the original on a subset of experiments.The authors report exact equivalence in that validation and manually checked reported numbers against raw outputs.Source: §D.5 · p. 17

E Extra Tables and Plots

  • Appendix E adds race/religion and gender bias tables, error-location comparisons, matched WinoBias error pairs, and annotated clean-prompt examples.The tables cover single error types and combinations of instruction-only, input-only, or both-part prompt errors.Source: Table E.1 · p. 18, Table E.2 · p. 18, Table E.3 · p. 19

Related papers

Loading 2610.29896v1…