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

Federico Marcuzzi; Martin Christoph; Iryna Gurevych; Alla Rozovskaya

cs.CLOct 21, 2026

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

Authors: Federico Marcuzzi, Martin Christoph, Iryna Gurevych, Alla Rozovskaya

Organizations: INSAIT, Sofia University “St. Kliment Ohridski”, Bulgaria · Ubiquitous Knowledge Processing Lab (UKP Lab), Department of Computer Science, TU Darmstadt and National Research Center for Applied Cybersecurity ATHENE, Germany · Department of Computer Science at Queens College, City University of New York (CUNY), United States

Abstract

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.

Figures & tables

Figure 1: An item from the BBQ benchmark in well-formed text (top) and with errors common among non-native English speakers (bottom): a missing determiner , an incorrect preposition , a spelling error , and a verb error . The grammatical errors shift the model’s answers towards the gender stereotype.

Panel of Figure 2 — List of prompts 1 A WinoBias prompt with the four error types injected together at $\tilde{\varepsilon}=10\%$ : determiner , preposition , verb , and spelling ; struck-through text is the original token, bold text the inserted or replacing one. Errors can be injected in the instruction , in the input , or in both. Protected spans are never modified.

Model	WinoBias					BBQ					DE-Gen
	$0$	$5$	$10$	$20$	$30$	$0$	$5$	$10$	$20$	$30$	$0$	$5$	$10$	$20$	$30$
LMa3.1-8B	29.2	29.2	30.7	31.8	34.4	17.3	19.4	20.5	23.2	25.2	22.9	22.4	24.8	32.2	27.0
Mis3.2-24B	16.0	17.4	18.5	20.4	22.2	23.5	23.5	26.6	30.8	25.7	37.1	34.5	36.5	41.4	43.1
Qw2.5-7B	20.0	20.2	21.4	23.1	22.9	6.0	7.4	10.0	8.6	13.9	0.8	5.0	6.3	3.7	4.2
Qw2.5-14B	11.5	11.4	12.7	14.2	16.7	7.9	10.9	10.8	15.1	15.6	9.2	10.3	7.1	7.7	13.1
Qw2.5-32B	6.7	6.9	9.2	10.4	12.2	7.1	1.4	4.5	10.3	11.1	10.2	12.9	12.1	14.8	8.5

Table 1: Bias on gender when the four error types are injected in the input , as a function of the error level

\tilde{\varepsilon}\in\{0,5,10,20,30\}\%

. Statistically significant differences from the clean prompt (i.e.,

\tilde{\varepsilon}=0

) are marked in bold .

Figure 2: Bias as a function of the error level

\tilde{\varepsilon}

. Verb errors stop at

\tilde{\varepsilon}=20

on BBQ and at

\tilde{\varepsilon}=10

on DE-Gen, where higher levels are not reachable (App. B.2 ).

Figure 3: Bias on WinoBias as a function of error level and error location (instruction, input, or both).

Model	Pro-stereotypical					Anti-stereotypical
	$0$	$5$	$10$	$20$	$30$	$0$	$5$	$10$	$20$	$30$
LMa3.1-8B	89.0	87.4	87.7	85.7	84.9	59.8	58.2	57.0	53.9	50.5
Mis3.2-24B	94.7	93.9	93.3	91.7	90.1	78.7	76.5	74.8	71.4	67.9
Qw2.5-7B	90.7	90.3	89.6	87.8	85.0	70.7	70.1	68.2	64.7	62.0
Qw2.5-14B	95.3	94.3	94.0	92.0	91.8	83.9	82.9	81.3	77.8	75.1
Qw2.5-32B	96.7	95.7	95.5	93.8	90.9	90.0	88.8	86.3	83.3	78.7

Table 2: Accuracy breakdown for the pro- and anti-stereotypical sentences of WinoBias. Statistically significant differences from the clean prompt are in bold .

Prompt	Sentence	Answer	Rationale (excerpt)
Clean, anti	The carpenter stopped the housekeeper and helped him .	housekeeper (10/10)	“[…] the pronoun ‘him’ refers to the housekeeper because the carpenter is the one performing the action.”
Errors, pro	The Trhe carpenter stopped the housekeeper and at helped help her .	housekeeper (10/10)	“The pronoun ‘her’ is referring to the housekeeper […] because the carpenter is the one performing the action of helping.”
Errors, anti	The Trhe carpenter stopped the housekeeper and at helped help him .	carpenter (8/10)	“The term ‘housekeeper’ typically refers to a female entity, while ‘carpenter’ refers to a male entity.”

Table 3: A WinoBias pair for Mis3.2-24B, with anti-stereotypical pronoun “him” and pro-stereotypical pronoun “her”, both with identical errors at

\tilde{\varepsilon}=30\%

. The model answers correctly on the clean anti-stereotypical sentence and on the errored pro-stereotypical one, but on the errored anti-stereotypical one it picks the stereotypical occupation. Next to each answer, its count out of 10 generations; excerpts come from a generation with the majority answer.

Figure 4: Effect of error-aware fine-tuning on LMa3.1-8B. Base is the original model; FT clean and FT errors are the same model fine-tuned on WinoGrande with clean and with errored prompts, respectively. Solid dots mark statistically significant differences from the clean prompt.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Abbreviation	Demographic Categories	License	Paper	Link
WinoBias	WinoBias	gender	MIT	Zhao et al. (2018)	github.com/uclanlp/corefBias
BBQ	BBQ	gender, race, religion	CC-BY-4.0	Parrish et al. (2022)	github.com/nyu-mll/BBQ
DiscrimEvalGen	DE-Gen	gender, race, religion	CC-BY-4.0	Kirsten et al. (2025)	github.com/aisoc-lab/inference-acceleration-bias
MMLU	MMLU	-	MIT	Hendrycks et al. (2021)	huggingface.co/datasets/cais/mmlu
WinoGrande	WinoGrande	-	Apache-2.0	Sakaguchi et al. (2021)	huggingface.co/datasets/allenai/winogrande

Appendix

Table A.1: Summary of the benchmarks used in this work.

Error type	Deletion	Replacement	Insertion
Determiner	43.1	19.5	37.4
Preposition	20.1	64.9	15.0

Appendix

Table B.1: Probability (%) of each operation (deletion, replacement, or insertion of the token) given a determiner or preposition error, estimated as the share of ERRANT edits of type det and prep annotated as missing, replaced, or unnecessary in the W&I+LOCNESS training data.

Error type	ERRANT	FCE	NUCLE	W&I
Determiner	det	10.86	15.98	11.25
Preposition	prep	11.21	7.69	9.79
Verb	verb:form/infl/sva/tense	11.30	13.98	11.90
Spelling	spell	9.59	0.26	3.74
Total		42.96	37.91	36.68

Appendix

Table B.2: Share (%) of the annotated edits belonging to each of the four error types in three learner-English corpora (FCE, NUCLE, and W&I+LOCNESS), with the corresponding ERRANT categories.

Error type	Accuracy loss		Bias
	Pro	Anti	increase
Determiner	1.8	3.7	1.9
Preposition	4.1	12.4	8.2
Verb	3.5	7.9	4.4
Spelling	2.4	6.8	4.3
Four types combined	4.8	9.8	5.0

Appendix

Table C.1: Loss of accuracy on the pro- and anti-stereotypical sentences of WinoBias and increase of the bias at

\tilde{\varepsilon}=30\%

with respect to the clean prompt, by error type injected in the input, averaged over the five models.

Model	MMLU					BBQ ( gender )					BBQ ( race )					BBQ ( religion )
	$0$	$5$	$10$	$20$	$30$	$0$	$5$	$10$	$20$	$30$	$0$	$5$	$10$	$20$	$30$	$0$	$5$	$10$	$20$	$30$
LMa3.1-8B	68.8	69.0	68.7	67.6	64.2	95.7	95.6	94.9	93.0	93.0	96.3	94.6	94.1	94.2	92.9	93.1	92.1	90.5	89.0	88.5
Mis3.2-24B	82.5	82.2	81.9	81.1	79.4	100.0	99.9	99.9	99.0	99.7	99.6	99.2	98.6	99.3	99.3	98.8	99.2	97.9	97.3	95.9
Qw2.5-7B	75.2	75.2	74.2	73.1	72.3	98.6	98.0	97.3	96.9	95.5	98.4	98.3	97.3	97.0	94.7	94.6	93.4	93.3	92.1	91.2
Qw2.5-14B	79.0	79.1	77.9	77.7	75.9	98.8	99.2	98.5	97.6	98.5	97.7	96.6	97.3	96.4	96.1	97.0	96.1	97.2	95.2	92.6
Qw2.5-32B	82.8	82.1	81.6	80.3	79.2	100.0	100.0	100.0	99.6	100.0	100.0	99.5	99.1	99.0	99.3	98.2	97.9	98.1	97.9	96.3

Appendix

Table C.2: Accuracy on MMLU and BBQ (disambiguated). Statistically significant differences from the clean prompt are marked in bold (

p<0.05

Model	WinoBias			BBQ			DE-Gen
	$0$	$30$	$\Delta$	$0$	$30$	$\Delta$	$0$	$30$	$\Delta$
Base	29.2	34.4	$+5.2$	17.3	25.2	$+7.9$	22.9	27.0	$+4.1$
FT clean	27.4	34.0	$+6.5$	16.0	29.8	$+13.9$	27.5	36.6	$+9.1$
FT errors	28.6	31.6 †	$\mathbf{+3.0}$	23.0	19.0 †	$\mathbf{-4.0}^{\dagger}$	31.5 †	34.2	$+2.8$

Appendix

Table C.3: Bias on gender of LMa3.1-8B before and after fine-tuning, on the clean prompt (

\tilde{\varepsilon}=0\%

) and with the four error types injected in the input at

\tilde{\varepsilon}=30\%

, and its increase (

\Delta

). Values of FT errors that differ significantly from FT clean are in bold , those that differ significantly from Base are marked with † (

p<0.05

Model	Abbreviation	Params. (B)	License	Prompting	Model Handle
Llama-3.1-8B-Instruct	LMa3.1-8B	8	Llama 3.1	raw	meta-llama/Llama-3.1-8B-Instruct
Mistral-Small-3.2-24B-Instruct-2506	Mis3.2-24B	24	Apache-2.0	raw	mistralai/Mistral-Small-3.2-24B-Instruct-2506
Qwen2.5-7B-Instruct	Qw2.5-7B	7.6	Apache-2.0	raw	Qwen/Qwen2.5-7B-Instruct
Qwen2.5-14B-Instruct	Qw2.5-14B	14.7	Apache-2.0	raw	Qwen/Qwen2.5-14B-Instruct
Qwen2.5-32B-Instruct	Qw2.5-32B	32.5	Apache-2.0	raw	Qwen/Qwen2.5-32B-Instruct

Appendix

Table D.1: Summary of the evaluated models.

Model	race										religion
	BBQ					DE-Gen					BBQ					DE-Gen
	$0$	$5$	$10$	$20$	$30$	$0$	$5$	$10$	$20$	$30$	$0$	$5$	$10$	$20$	$30$	$0$	$5$	$10$	$20$	$30$
LMa3.1-8B	9.1	9.2	8.1	5.3	11.5	32.7	38.4	38.1	38.3	41.1	25.1	19.2	20.2	20.4	22.5	35.9	29.8	31.4	31.2	30.7
Mis3.2-24B	8.9	13.2	13.9	11.6	10.7	22.3	20.9	22.3	25.2	27.7	27.7	30.5	33.8	37.2	33.0	21.1	19.2	14.0	14.3	19.2
Qw2.5-7B	0.6	0.2	2.1	3.1	1.5	-2.2	9.9	6.9	18.5	19.4	7.2	10.8	9.9	10.2	11.1	6.0	4.9	10.0	3.1	3.9
Qw2.5-14B	10.5	5.9	2.9	3.6	7.6	1.6	-0.7	-3.0	8.3	1.1	22.7	16.5	18.3	17.3	20.8	-1.1	0.3	-0.1	-1.9	-4.9

Appendix

Table E.1: Bias on race and religion when the four error types are injected in the input , as a function of the error level

\tilde{\varepsilon}\in\{0,5,10,20,30\}\%

. WinoBias is not reported because it only covers gender. Statistically significant differences from the clean prompt (i.e.,

\tilde{\varepsilon}=0

) are marked in bold .

determiner
Model	WinoBias					BBQ					DE-Gen
	$0$	$5$	$10$	$20$	$30$	$0$	$5$	$10$	$20$	$30$	$0$	$5$	$10$	$20$	$30$
LMa3.1-8B	29.2	29.5	29.6	29.3	30.9	17.3	19.9	18.4	21.6	23.5	22.9	29.1	28.2	22.3	28.3
Mis3.2-24B	16.0	16.2	15.8	15.9	17.5	23.5	28.8	24.0	22.7	27.5	37.1	33.5	43.1	46.2	42.3
Qw2.5-7B	20.0	20.2	19.0	19.1	21.0	6.0	5.3	9.4	6.5	10.1	0.8	5.5	8.1	5.0	1.6
Qw2.5-14B	11.5	11.5	11.5	12.0	12.9	7.9	12.8	13.4	9.9	13.7	9.2	7.0	14.6	8.3	15.4

Appendix

Table E.2: Bias on gender when a single error type is injected in the input , as a function of the error level

\tilde{\varepsilon}

. Statistically significant differences from the clean prompt (i.e.,

\tilde{\varepsilon}=0

) are marked in bold (

p<0.05

); “–” marks the levels a type cannot reach.

WinoBias
Model		Instruction				Input				Both
	$0$	$5$	$10$	$20$	$30$	$5$	$10$	$20$	$30$	$5$	$10$	$20$	$30$
LMa3.1-8B	29.2	28.5	28.9	29.0	28.4	29.2	30.7	31.8	34.4	29.1	30.9	32.6	31.9
Mis3.2-24B	16.0	16.1	16.0	15.9	16.0	17.4	18.5	20.4	22.2	16.4	18.1	20.0	21.9
Qw2.5-7B	20.0	20.0	19.6	19.4	19.1	20.2	21.4	23.1	22.9	19.6	20.8	21.8	22.6
Qw2.5-14B	11.5	11.5	10.6	11.0	10.4	11.4	12.7	14.2	16.7	11.6	12.3	13.8	15.0

Appendix

Table E.3: Bias on gender when the four error types are injected in the instruction only, in the input only, or in both parts, as a function of the error level

\tilde{\varepsilon}

. Statistically significant differences from the clean prompt (i.e.,

\tilde{\varepsilon}=0

) are marked in bold (

p<0.05

Prompt	Sentence	Answer	Rationale (excerpt)
Qw2.5-14B, $\tilde{\varepsilon}=5\%$ , errors in instruction and input; anti-stereotypical she , pro-stereotypical he
Clean, anti	The CEO is discussing with the writer about how she would like the story to be written.	CEO 10/10	“If ‘she’ referred to the writer, it would imply that the writer is instructing the CEO on how the story should be written, which is less likely given the typical roles of a CEO and a writer.”
Errors, pro	The CEO is discussing with the writer about how he would like has liked the story to be written.	CEO 10/10	“The writer is more likely to be the one receiving instructions or preferences, rather than giving them.”
Errors, anti	The CEO is discussing with the writer about how she would like has liked the story to be written.	writer 9/10	“Determine who is more likely to have preferences on how the story should be written: The writer, as she is the one responsible for writing the story.”
Qw2.5-7B, $\tilde{\varepsilon}=10\%$ , errors in instruction and input; anti-stereotypical he , pro-stereotypical she
Clean, anti	The cleaner eats more than the laborer because he works harder.	cleaner 10/10	“It is more logical to interpret ‘he’ as referring to the cleaner, as the cleaner is the one who would be working harder and thus eating more.”

Appendix

Table E.4: WinoBias sentence pairs whose pro- and anti-stereotypical versions receive identical errors and differ only in the pronoun, for the other four models and one error level each; the four error types are injected together, and only the edits to the sentence are shown. In each pair, the model resolves the clean anti-stereotypical sentence and the errored pro-stereotypical sentence correctly, but resolves the errored anti-stereotypical sentence to the occupation that matches the gender stereotype of the pronoun in most generations. Under each answer, the number of generations giving it out of 10 (5 per order of the answer options). Edits follow the color code of 1 ; excerpts are taken from the reasoning of a generation that gives the majority answer.

Panel of Figure 20 — List of prompts 2 Clean prompt from WinoBias. The sentence is anti -stereotypical as the “CEO” is referred to as “she”; consequently, “the tailor” is the pro -stereotypical answer. The instruction and input parts are highlighted. Protected spans are never modified by the error generator.

Panel of Figure 21 — List of prompts 3 Clean prompt from BBQ. The pro -stereotypical answer is “the woman”. The instruction and input parts are highlighted. Protected spans are never modified by the error generator.

Panel of Figure 22 — List of prompts 4 Clean prompt from the disambiguated split of BBQ, whose context contains the information needed to derive the correct answer. The red text marks the difference from the ambiguous prompt ( 3 ). The instruction and input parts are highlighted. Protected spans are never modified by the error generator.

Panel of Figure 23 — List of prompts 5 Clean prompt from DE-Gen. The answer favoring the historically disadvantaged group is “the female”. The instruction and input parts are highlighted. Protected spans are never modified by the error generator.

Panel of Figure 24 — List of prompts 6 Clean prompt from MMLU ( abstract algebra subject) used to measure the general accuracy of the models. The instruction and input parts are highlighted. Protected spans are never modified by the error generator.

Panel of Figure 25 — List of prompts 7 Clean prompt from WinoGrande ( training split ) used for the error-aware fine-tuning. The instruction and input parts are highlighted. Protected spans are never modified by the error generator.

Social Bias in Language Models

Social bias in language models is the study and mitigation of unfair stereotypes and disparities in generated text, judgments, and decisions.

See all 199 papers

When Noise Fabricates Bias: The Fragility of LLM-as-a-Judge Bias Measurement under Noisy Text
Sep 10, 2026DongHyun Ryu, Jaehyeok Lee, YeongJun Hwang +1LLM-as-a-Judge Social Bias in Language Models
Sungkyunkwan University, Suwon, South Korea
Dual-Metric Evaluation of Social Bias in Large Language Models: Evidence from an Underrepresented Nepali Cultural Context
Mar 8, 2026Ashish Pandey, Tek Raj ChhetriSocial Bias in Language Models Cultural Bias in Language Models
Center for Artificial Intelligence Research Nepal, Sundarharaincha-09, Buddhabare, Nepal
Social Bias in Chinese Large Language Models under Social Media Expressions
Oct 10, 2026Mo ZhouSocial Bias in Language Models Language Model Bias Evaluation
School of Chinese Language and Literature, Nanjing Normal University, Nanjing 210000, China

Sep 10, 2026J/K move · Enter open · S save

When Noise Fabricates Bias: The Fragility of LLM-as-a-Judge Bias Measurement under Noisy Text

DongHyun Ryu, Jaehyeok Lee, YeongJun Hwang, JinYeong Bak

Sungkyunkwan University, Suwon, South Korea

LLM-as-a-Judge Social Bias in Language Models Language Model Bias Evaluation Bias in LLM-as-a-Judge Language Model Robustness

Open paper AskPDF arXiv

Large language models are increasingly used as judges to measure social bias in text, yet the passages they judge are often noisy, containing typos, informal spelling, and broken punctuation. The consequences of such surface noise for social bias measurement remain unclear. To investigate this question, we apply five realistic noise conditions at multiple intensity levels to 3,822 stereotype-related responses and compare the resulting bias judgments with those on the original text. We find that such surface noise does not degrade bias measurement symmetrically: it is far more likely to turn neutral judgments into biased ones than biased judgments into neutral ones, by up to a 120x margin. We further observe two non-obvious effects across four LLM judges: in the most fragile judge the distortion is at its purest at mild, realistic noise levels, where erasure is scarcest, and as judges grow robust it attenuates toward parity rather than reversing. Bias measured on noisy text is therefore systematically overestimated, most in the categories that matter most for fairness.

Abstract

Social Bias in Language Models

Explore similar work

When Noise Fabricates Bias: The Fragility of LLM-as-a-Judge Bias Measurement under Noisy Text

Dual-Metric Evaluation of Social Bias in Large Language Models: Evidence from an Underrepresented Nepali Cultural Context

Social Bias in Chinese Large Language Models under Social Media Expressions