Bias and Fairness in Large Language Models: A Survey

• LLMs can perpetuate and amplify stereotypes, derogatory language, exclusion, toxicity, and dialect-related harms affecting marginalized communities (Bender et al. 2021; Dodge et al. 2021; Sheng et al. 2021b).These harms can reproduce injustice and reinforce systems of inequity.Source: §1 · p. 2
• It organizes bias-evaluation metrics by model access and operation level: embeddings, token probabilities, and generated text.The survey separates metrics from datasets and formalizes metrics with unified notation while discussing limitations for downstream application bias.Source: §1 · pp. 2–3
• It classifies mitigation methods by intervention stage—pre-processing, in-training, intra-processing, and post-processing—and analyzes method similarities, trends, and persistent bias.Several techniques receive mathematical formalization and representative examples.Source: §1 · pp. 2–3
• The survey formalizes social bias and fairness for NLP by distinguishing social harms and introducing fairness desiderata for LLMs.Its taxonomy is intended to improve precision when describing bias evaluation and mitigation efforts.Source: §1 · pp. 2–3
• It categorizes evaluation datasets as counterfactual inputs or prompts, identifies targeted harms and social groups, and compiles publicly available datasets.The taxonomy also examines compatibility with metrics and highlights consistency, reliability, and validity challenges.Source: §1 · p. 3
• The survey covers LLM bias research for English, with additional machine-translation and multilingual work, focusing on studies proposing novel metrics, datasets, or mitigation techniques.It excludes general algorithmic fairness and broad language-technology bias surveys, while identifying challenges involving power imbalances, evaluation standards, mitigation, and fairness guarantees.Source: §1 · pp. 3–4

2.1 Preliminaries

• LLMs include autoregressive, autoencoding, and encoder-decoder architectures trained on hundreds of millions to trillions of tokens, then adapted across tasks and value-constrained applications.The survey focuses especially on task-agnostic adaptations that constrain model behavior for bias mitigation and reduction.Source: §2.1 · pp. 4–5

2.2 Defining Bias for LLMs

• Bias and fairness definitions must identify who is harmed, why, and how harms reflect social hierarchies, because imprecise or inconsistent definitions obscure which injustices technical methods address.The survey treats bias broadly, recognizes its normative and context-dependent character, and notes that “debiased” does not imply complete bias removal.Source: §2.2.1 · pp. 5–6

2.2.1 Social Bias and Fairness

• Social bias encompasses representational and allocational harms arising from historical and structural power asymmetries, which may overlap and manifest differently across social groups and intersectional identities.The taxonomy includes harms such as stereotyping, misrepresentation, derogatory language, exclusionary norms, and discrimination.Source: §2.2.1 · pp. 6–7, Table 1 · p. 8

2.2.2 Bias in NLP Tasks

• Bias can arise across generation, translation, retrieval, question-answering, inference, and classification, where outputs may encode stereotypes, exclusionary norms, or disparate performance.These task-specific outputs can often be evaluated through scores conditioned on social-group membership, such as token probabilities, sentiment, or gendered-language indicators.Source: §2.2.2 · pp. 8–9

2.2.3 Bias in the Development and Deployment Life Cycle

• Bias may emerge during training data construction, modeling, evaluation, or deployment through unrepresentative data, optimization choices, benchmark limitations, metric selection, or changed usage settings.Historical and structural biases can persist even in properly collected data, while decoding, ranking, and deployment interfaces can further affect model behavior.Source: §2.2.3 · pp. 9–10

2.3 Fairness Desiderata for LLMs

• Group, subgroup, and individual fairness provide useful frameworks but do not determine a universal fairness constraint; specifications remain subjective, contextual, value-dependent, and stakeholder-dependent.Group parity can still conceal unfairness for intersectional subgroups, motivating attention to overlapping population subsets and individual similarity.Source: §2.2.1 · p. 7, §2.3 · p. 10
• The survey proposes LLM fairness desiderata including unawareness, invariance, equal social-group associations, equal neutral associations, and replicated distributions.These desiderata generalize outcome- and error-disparity ideas beyond conventional classification, especially for ambiguous language-generation settings.Source: §2.3 · pp. 10–11

2.4 Overview of Taxonomies

• The overview organizes bias-evaluation metrics by model access and dataset structure, datasets by counterfactual inputs or prompts, and evaluation methods by the harms and groups they target.Metric levels include embeddings, token probabilities, and generated text; dataset formats include masked or unmasked counterfactual sentences, sentence completions, and question-answering prompts.Source: §2.4 · pp. 11–12
• The mitigation taxonomy organizes techniques by their intervention stage in the LLM workflow: pre-processing, in-training, intra-processing, and post-processing.The complete taxonomy is described in Section 5.Source: §2.4 · p. 12
• Pre-processing methods change model inputs through data augmentation, filtering or reweighting, data generation, instruction tuning, or projection-based mitigation.These methods respectively extend, remove or reweight, generate, prepend tokens, or transform hidden representations.Source: §2.4 · p. 12
• In-training methods modify model parameters via gradient-based updates, including architecture, loss-function, selective-parameter, and parameter-filtering changes.These subclasses alter model configuration, objectives, updated parameter subsets, or retained parameter subsets.Source: §2.4 · p. 12
• Intra-processing methods modify inference behavior without further training through decoding-strategy changes, weight redistribution, or modular debiasing networks.These techniques modify probabilities, attention-weight entropy, or add stand-alone components.Source: §2.4 · p. 12
• Post-processing methods modify the output text generations produced by the model.This stage intervenes after generation rather than changing inputs, parameters, or inference behavior.Source: §2.4 · p. 12

3. Taxonomy of Metrics for Bias Evaluation

• The taxonomy organizes fairness metrics by task, bias type, dataset structure, and model input, including embeddings, probabilities, and generated text.The authors formalize the metrics, provide examples, and discuss their challenges and limitations.Source: §3 · p. 12, §3.1 · p. 13

3.1 Facets of Evaluation of Biases: Metrics and Datasets

• Gender substitution changes the question-answering response, associating male data engineers with data sources and female data engineers with technical skills.The example implies that male data engineers are assumed to possess technical skills that female data engineers are advised to develop.Source: Figure 1 · p. 13, §3.1 · p. 13
• In text completion, substituting gender produces the same continuation, showing that bias can be explicit in one task and implicit in another.The question-answering example changes across genders, whereas the completion example does not.Source: Figure 1 · p. 13, §3.1 · p. 13
• Bias metrics and datasets are often task-specific because text generation, classification, and question-answering manifest biases differently.The paper uses two tasks in Figure 1 to illustrate this task dependence.Source: §3.1 · p. 13, Figure 1 · p. 13
• The taxonomy distinguishes bias evaluation by dataset structure, including paired sentences where one sentence is biased and the other is less biased.CrowS-Pairs exemplifies stereotype sentence pairs, while alternative pseudo-log-likelihood metrics can also operate on that dataset.Source: §3.1 · p. 13
• The evaluation framework treats metrics and datasets as distinct, because one dataset can support multiple metrics rather than defining a metric by its original dataset.The paper recommends decomposing datasets from their originally paired metrics and identifying alternatives for each dataset class.Source: §3.1 · pp. 13–14
• For a dataset D, the appropriate metrics form a subset ψ(D) of all metrics, determined largely by the dataset’s structure and task.The paper highlights this relationship because many datasets share structures such as sentence pairs.Source: §3.1 · p. 14, Figure 2 · p. 14

Taxonomy of Metrics based on What They Use

• The taxonomy groups bias metrics into embedding-based, probability-based, and generated text-based categories according to their model inputs.Embedding metrics use dense vector representations, probability metrics use model-assigned probabilities, and generated-text metrics analyze outputs conditioned on prompts.Source: §3.2 · p. 14

3.3 Embedding-Based Metrics

• Embedding-based metrics compare neutral words, such as professions, with identity-related words, such as gender pronouns, using vector-space distances and cosine similarity.The section focuses on one static word-embedding method and primarily on sentence-level contextualized embeddings for LLMs.Source: §3.3 · pp. 14–15, Figure 3 · p. 17

3.3.1 Word Embedding Metrics

• WEAT measures associations between protected social-group concepts and neutral attributes through cosine-based similarity statistics, with larger effect sizes indicating stronger bias.WEAT (Caliskan, Bryson, and Narayanan 2017) emulates the Implicit Association Test, while WEAT* (Dev et al. 2021) uses definitionally gendered words to capture stronger associations.Source: §3.3.1 · pp. 15, 17

3.3.2 Sentence Embedding Metrics

• Sentence-level methods adapt WEAT to contextualized embeddings through templates, sampling, or word-level aggregation, enabling more targeted evaluation of bias dimensions.SEAT (May et al. 2019) uses template sentences and [CLS] embeddings; CEAT (Guo and Caliskan 2021) estimates bias from a random-effects distribution; Sentence Bias Score (Dolci, Azzalini, and Tanelli 2023) weights word-level gender-direction bias by semantic importance.Source: §3.3.2 · pp. 17–18

3.3.3 Discussion and Limitations

• Embedding-space bias has weak or inconsistent relationships with downstream task bias, and several studies caution that representation bias should not be conflated with application disparities.Goldfarb-Tarrant et al. (2021) find no reliable correlation, while Cabello, Jørgensen, and Søgaard (2023) show such associations can be independent of downstream disparities under violated assumptions.Source: §3.3.3 · p. 18
• Embedding-based measurements are highly sensitive to templates, seed words, representation choices, and whether they capture subtle harms beyond word associations.Gonen and Goldberg (2019) show debiasing may represent bias in new ways, while Delobelle et al. (2022) recommend task-specific downstream metrics; coverage of derogatory language, toxicity, exclusionary norms, and disparate performance remains unclear.Source: §3.3.3 · pp. 18–19

3.4 Probability-Based Metrics

• Probability-based metrics compare predicted token probabilities across inputs with perturbed protected attributes, operating on template sentences or sentence sets.The section illustrates these techniques in Figure 4.Source: §3.4 · p. 19

3.4.1 Masked Token Methods.

• Masked-token methods use fill-in-the-blank predictions, including DisCo’s differing top-three completions and LPBS’s prior-normalized probabilities for neutral attributes.LPBS measures bias attributable to the neutral attribute by comparing normalized probabilities for opposing social-group words; Categorical Bias Score extends normalized log probabilities to non-binary targets.Source: §3.4.1 · pp. 19–20

3.4.2 Pseudo-Log-Likelihood Methods.

• Pseudo-log-likelihood methods mask tokens sequentially to estimate sentence probabilities, supporting comparisons such as CrowS-Pairs’ stereotypical versus less-stereotypical preferences and CAT’s stereotype, anti-stereotype, and meaningless options.CAT evaluates P(M|U, θ), whereas CrowS-Pairs evaluates P(U|M, θ).Source: §3.4.2 · p. 20
• AUL predicts all tokens from unmasked sentences and supports multiple correct candidates, while AULA additionally weights predictions by token attention.These variants aim to improve prediction accuracy and avoid masking-selection bias; an ideal average bias score for the described comparisons is 0.5.Source: §3.4.2 · p. 21

3.4.3 Discussion and Limitations.

• Probability-based metrics may correlate weakly with downstream-task bias and therefore should be paired with metrics that directly assess the downstream task.Masked-token templates may lack semantic and syntactic diversity, while stereotype comparisons can reduce behavior to sentence rankings and rely on contested equal-selection assumptions.Source: §3.4.3 · pp. 21–22
• Most metrics assume binary groups or pairs, and equal word predictions may conflate benign linguistic associations with harmful stereotypes, making explicit fairness criteria necessary.These assumptions may fail to represent complex group relationships or distinguish non-stereotypical knowledge from bias.Source: §3.4.3 · p. 22

3.5 Generated Text-Based Metrics

• The standard evaluation conditions an LLM on bias- or toxicity-associated prompts and assesses its continuation, enabling direct output analysis when probabilities or embeddings are unavailable.Common resources include RealToxicityPrompts (Gehman et al. 2020) and BOLD (Dhamala et al. 2021), alongside templates with perturbed social groups.Source: §3.5 · p. 22

3.5.1 Distribution Metrics.

• Distribution metrics compare associations between neutral words and demographic terms, or social-group token distributions, against reference distributions such as uniform, and measure demographic representation and stereotypical associations.Source: §3.5.1 · pp. 22–23, Figure 5 · p. 23

3.5.2 Classifier Metrics.

• Classifier metrics detect group-conditioned differences by scoring generated text for toxicity, sentiment, regard, style, or other targeted bias dimensions.Examples include Expected Maximum Toxicity, Toxicity Probability, Toxic Fraction, Score Parity (Sicilia and Alikhani 2023), Counterfactual Sentiment Bias (Huang et al. 2020), Regard Score (Sheng et al. 2019), and Full Gen Bias (Smith et al. 2022).Source: §3.5.2 · pp. 24–25

3.5.3 Lexicon Metrics.

• Lexicon metrics analyze words individually using harmful-word lists or precomputed affective and gender-bias scores, including HONEST and BOLD-associated measures.HONEST (Nozza, Bianchi, and Hovy 2021) counts hurtful completions, while Psycholinguistic Norms and Gender Polarity (Dhamala et al. 2021) aggregate lexicon values.Source: Figure 5 · p. 23, §3.5.3 · pp. 25–26

3.5.4 Discussion and Limitations.

• Decoding parameters can drastically change generated-text bias scores, produce contradictory results, and vary across datasets, so prompts and settings should be reported.Akyürek et al. (2022) identify generation length, sampling temperature, and beam-search top-k as influential choices.Source: §3.5.4 · p. 26
• Generated-text metrics have important validity limits: associations may poorly proxy disparities, classifiers may be biased or unreliable, and lexicons may miss relational harms.Co-occurrence can confuse mention with use, toxicity and sentiment classifiers can misclassify protected groups, and harmless individual words can form biased sequences.Source: §3.5.4 · p. 26

• Embedding- and probability-based metrics should not be the sole bias measures because their relationships with downstream biases can be weak and unreliable.They should be accompanied by direct evaluation of the downstream task.Source: §3.6 · p. 27
• Bias evaluations should report model hyperparameters, templates, and prompts, which can produce contradictory conclusions about a model’s bias.Source: §3.6 · p. 27
• New metrics should reflect targeted harms and power dynamics, use valid auxiliary resources, and satisfy Cao et al. (2022b)’s desiderata for theory-grounded, natural, specific, exhaustive, and generalizable evaluations.These desiderata include generalizing to unconsidered groups, grounding in social science theory, covering possible biases, using natural text inputs, and evaluating specific instances rather than abstractions.Source: §3.6 · p. 27

4. Taxonomy of Datasets for Bias Evaluation

• The taxonomy distinguishes counterfactual inputs from prompts and summarizes each dataset’s instances, targeted bias issues, and social groups in Table 4.The authors also compile publicly available datasets to facilitate access and use.Source: §4 · p. 27, Table 4 · p. 28

4.1 Counterfactual Inputs

• Counterfactual datasets perturb social-group terms while preserving meaning, enabling bias detection through changes in token probabilities, sentence likelihoods, or generated continuations.They comprise masked-token and unmasked-sentence formats, which can be adapted to one another and paired with corresponding metrics.Source: §4.1 · pp. 27–29, §4.1.2 · p. 30

4.1.1 Masked Tokens.

• Masked-token datasets use fill-in-the-blank options for coreference, stereotypes, and other biases, supporting probability, pseudo-log-likelihood, and accuracy-based evaluation.Examples include WinoBias, Winogender, GAP, BUG, StereoSet, and BEC-Pro, which target gender, race, religion, profession, and related associations.Source: §4.1.1 · pp. 29–30

4.1.2 Unmasked Sentences.

• Unmasked-sentence datasets compare paired sentences differing in demographic terms, covering stereotypes, sentiment, conversational harms, identity associations, and robustness to demographic perturbations.Examples include CrowS-Pairs, Equity Evaluation Corpus, RedditBias, HolisticBias, WinoQueer, Bias-STS-B, PANDA, and Bias NLI.Source: §4.1.2 · pp. 30–31

4.1.3 Discussion and Limitations.

• Counterfactual benchmarks face construct-validity problems: nearly half of instances in several datasets contain ambiguous stereotypes, unclear harms, or invalid social-group perturbations.Small semantic-preserving edits can also drastically change Winogender and Bias NLI scores, while U.S.-specific, template-based data may not generalize broadly.Source: §4.1.3 · pp. 31–32

4.2 Prompts

• Prompt datasets elicit continuations or answers and are therefore suited to generated-text metrics, but social-group labels can create ambiguity about whose bias or harm the output represents.Situation-based prompts that inspect generated social-group identifiers are proposed as a less misleading alternative.Source: §4.2 · p. 32

4.2.1 Sentence Completions.

• Sentence-completion datasets use naturalistic prefixes or cloze prompts to assess toxicity, profession, gender, race, religion, ideology, and multilingual gender stereotypes in generated text.RealToxicityPrompts provides 100,000 prefixes, BOLD provides 23,679 prompts, and HONEST provides 420 sentences across six languages.Source: §4.2.1 · pp. 32–33

4.2.2 Question-Answering.

• Question-answering and retrieval datasets probe stereotype reliance or representation disparities through ambiguous and disambiguated questions, underspecified alternatives, and gender-varied document retrieval.BBQ contains 58,492 examples across nine social groups, while Grep-BiasIR uses 118 queries and 708 documents.Source: §4.2.2 · p. 33

Taxonomy overview

• The taxonomy distinguishes mitigation by intervention stage: pre-processing changes data or prompts, in-training changes training, intra-processing changes an accessible trained model without further training, and post-processing changes outputs from a black-box model.Table 5 summarizes the techniques, while Figure 6 illustrates each stage’s pathways, inputs, and outputs.Source: §5 · p. 34, Table 5 · p. 34, Figure 6 · p. 35

5.1 Pre-Processing Mitigation

• Pre-processing methods transform data, prompts, or representations before fine-tuning, including augmentation, filtering, reweighting, generation, instruction tuning, and projection.A pre-trained model can be fine-tuned on transformed data and prompts or initialized with transformed representations.Source: §5.1 · p. 35

5.1.1 Data Augmentation

• Data augmentation expands underrepresented groups through counterfactual swaps, selective replacement, interpolation, or added examples, but may cause grammatical errors and depend on incomplete word-pair lists.Counterfactual data augmentation can use one-sided or two-sided examples, while selective methods target efficient or equity-promoting counterfactuals.Source: §5.1.1 · pp. 35–37, §5.1.2 · p. 37

5.1.2 Data Filtering and Reweighting

• Data filtering and reweighting select, remove, mask, or downweight examples and can also equalize teacher probabilities to reduce demographic shortcuts during training.Methods target underrepresented or low-bias data, highly biased examples, protected-attribute proxies, or instances with disproportionate influence on unfairness.Source: §5.1.2 · pp. 37–38

5.1.3 Data Generation

• Data generation creates standards-targeted datasets, human-written exemplary dialogues, recovery responses, prosocial responses, and broader word lists when existing examples or lists are insufficient.Crowdsourced generation can encode majority or culturally narrow norms, while warmth-and-competence word lists broaden beyond group-specific stereotypes.Source: §5.1.3 · pp. 38–39

5.1.4 Instruction Tuning

• Instruction tuning modifies prompts with textual triggers, control or reward tokens, or trainable continuous prefixes to condition less biased generations without initially changing model parameters.Modified language and control tokens are interpretable, whereas continuous prompt tuning freezes original parameters and learns virtual task-specific tokens.Source: §5.1.4 · pp. 39–40, §5.1.6 · p. 41

5.1.5 Projection-based Mitigation

• Projection methods remove or separate protected-attribute directions from contextualized embeddings before fine-tuning, while less-aggressive approaches preserve grammatical and semantic information.INLP, IGBP, and Sent-Debias project away bias subspaces; OSCAR and related methods retain relevant gendered meaning, though embedding bias weakly predicts downstream bias.Source: §5.1.5 · p. 40, §5.1.6 · p. 41

5.1.6 Discussion and Limitations

• Pre-processing can be limited by binary social assumptions, proxy-based or nonequivalent word lists, distributional distortions, ineffective prompts, majority norms, and weak links between embedding and downstream bias.The section suggests cost-sensitive reweighting, careful dataset curation, and documenting authors’ identities, backgrounds, and perspectives as stronger future directions.Source: §5.1.6 · p. 41

5.2 In-Training Mitigation

• In-training mitigations modify trainable model parameters through architecture changes, loss functions, selective updates, or parameter filtering.These approaches alter optimization, next-word probabilities, fine-tuning parameter selection, or specific neurons and weights.Source: §5.2 · p. 41, Figure 8 · p. 42

5.2.1 Architecture Modification.

• Architecture modifications add debiasing components such as ADELE adapters, freeze pre-trained layers during fine-tuning, or combine shared and demographic-specific encoders.ADELE learns debiasing knowledge from the BEC-Pro gender-bias dataset while updating only injected adapter layers.Source: §5.2.1 · p. 42

5.2.2 Loss Function Modification.

• Loss-function modifications impose embedding, attention, predicted-distribution, contrastive, adversarial, or reinforcement-learning objectives to reduce associations between social groups and biased outputs.Fairness objectives are commonly added as regularization to preserve pretrained knowledge, while methods may equalize counterfactual representations or probabilities, redistribute attention, hide protected attributes, or reward less biased generations.Source: §5.2.2 · pp. 42, 45, 46, 47, 48

5.2.3 Selective Parameter Updating.

• Selective parameter updating freezes most pretrained weights to reduce catastrophic forgetting and computational expense while tuning targeted parameters during fine-tuning.Examples freeze over 99% of parameters, update only attention matrices, or select weights with the greatest domain-specific bias contributions using contrastive-pair gradients.Source: §5.2.3 · p. 49

5.2.4 Filtering Model Parameters.

• Pruning 10–20% of model parameters increases resistance to jailbreaking, whereas pruning 30% or more can increase gender, race, and religious bias, with effects depending on pruning degree.More extensive pruning can have detrimental effects, and pruning may amplify bias in some stereotyping and toxicity classification settings.Source: §5.2.4 · pp. 49–50

5.2.5 Discussion and Limitations.

• In-training mitigation requires a trainable model, can be computationally expensive, and may impair language understanding through catastrophic forgetting; effectiveness depends on the targeted mechanism and fairness assumptions.Embedding-based objectives may be limited by weak links between embedding and downstream-task bias, whereas attention-based objectives may be more effective; empirical comparisons remain limited and fairness definitions should be made explicit.Source: §5.2.5 · p. 50

5.3 Intra-Processing Mitigation

• Intra-processing methods modify pretrained models at inference to generate debiased predictions without further training or fine-tuning, including decoding, parameter, and modular-network interventions.These techniques are inference-stage mitigations that change model behavior while preserving the underlying training process.Source: §5.3 · p. 50

5.3.1 Decoding Strategy Modification.

• Decoding strategies constrain token selection, rerank outputs, modify token probabilities, compare expert and anti-expert distributions, or project embeddings to reduce biased generation.Examples include offensive-token blocking, counterfactual beam search, toxicity-guided logits, self-debiasing, and nullspace projection for gender- or religion-invariant tokens.Source: §5.3.1 · pp. 51–52

5.3.2 Weight Redistribution.

• Post hoc weight redistribution changes attention entropy through temperature scaling, broadening focus toward potentially less stereotypical tokens or narrowing context to reduce stereotypical exposure.The scaling hyperparameter can be tuned to maximize a fairness metric without additional training.Source: §5.3.2 · p. 53

5.3.3 Modular Debiasing Networks.

• Modular debiasing networks store bias-specific parameter changes in subnetworks or adapters that can be selectively fused with a base pretrained model during inference.These modules support different protected attributes or debiasing tasks while preserving the original model for contexts where protected-attribute information is useful.Source: §5.3.3 · p. 53

5.3.4 Discussion and Limitations.

• Intra-processing research is concentrated on decoding, whose toxicity filters can suppress minority dialects, amplify bias, or generate harmful language when their classifiers or heuristics are biased.Weight redistribution and modular debiasing remain comparatively limited, while decoding must balance mitigation against diverse output generation.Source: §5.3.4 · pp. 53–54

5.4 Post-Processing Mitigation

• Post-processing methods mitigate bias in generated outputs without accessing or changing the original model parameters, with examples organized around biased-token identification and replacement.These methods address settings where models expose outputs but provide limited information about training data, optimization, or internal structure.Source: §5.4 · p. 54

5.4.1 Rewriting.

• Rewriting strategies replace harmful or biased language in fully generated outputs using keyword replacement, machine translation, parallel corpora, or neural rewriters.Approaches include LIME- or SHAP-guided replacement, protected-attribute masking, gender-neutral translation, politeness and subjectivity correction, agency editing, and user-informed modification (Tokpo and Calders 2022; Dhingra et al. 2023; He, Majumder, and McAuley 2021; Jain et al. 2021; Sun et al. 2021; Vanmassenhove, Emmery, and Shterionov 2021; Amrhein et al. 2023; Wang et al. 2022; Pryzant et al. 2020; Ma et al. 2020; Majumder, He, and McAuley 2022).Source: §5.4.1 · pp. 54–55

5.4.2 Discussion and Limitations.

• Rewriting can be biased because deciding which outputs to alter is subjective, while protected-attribute removal may erase context and produce less diverse outputs.Systems should avoid disproportionately flagging particular groups’ language styles when applying toxicity- or sentiment-like detection criteria.Source: §5.4.2 · p. 55
• Neural rewriting methods are constrained by the availability of parallel training corpora, limiting the dimensions of bias they can address.This limitation compounds the risk that rewriting methods reproduce biases in their training data or intervention criteria.Source: §5.4.2 · p. 55

Recommendations

• Mitigation methods should avoid flattening power imbalances by validating that identity-word lists accurately and completely represent the social groups they model.Masking or replacing identity words may miss power dynamics specific to particular groups.Source: §5.5 · p. 55
• Fairness objectives should state their encoded assumptions, align with explicit fairness desiderata, and consider cost-sensitive learning to weight minority classes more heavily.Source: §5.5 · pp. 55–56
• Bias mitigation should preserve output diversity by validating that toxicity heuristics do not marginalize minoritized groups or their dialects and usages.Source: §5.5 · p. 56
• Post-processing should preserve important contexts by recognizing the value-laden nature of rewriting and avoiding flattened styles, variation, or erased social identities.Source: §5.5 · p. 56

6.1 Addressing Power Imbalances

• Future LLM research should center marginalized communities through participatory development, culturally diverse values and languages, and data governance that preserves community agency.This includes expanding beyond Western and English-centric assumptions, documenting resources, protecting privacy and consent, and sharing benefits with data-originating communities.Source: §6.1 · pp. 56–57

6.2 Conceptualizing Fairness for NLP

• Fairness research should define context-specific desiderata, disaggregate social groups, recognize distinct oppression mechanisms, and include neglected, intersectional, and unobserved identities.Definitions should make normative and historical contexts explicit rather than treating sameness, binary categories, or protected-attribute erasure as universally appropriate.Source: §6.2 · pp. 57–58

6.3 Refining Evaluation Principles

• Evaluation needs reporting standards, carefully contextualized benchmarks, reliable and valid or living datasets, alternative testing paradigms, and broader metrics beyond toxicity and group fairness.Comprehensive benchmarks may improve comparability but can obscure contextual harms, encode dominant-group assumptions, and foster complacency when underlying power imbalances remain.Source: §6.3 · pp. 58–59

6.4 Improving Mitigation Efforts

• Mitigation research should scale scarce resources, compare intervention stages, develop hybrid techniques, and investigate how LLM components encode bias and respond to mitigation.Bias reduced in embedding space can re-emerge downstream, motivating empirical comparisons across pre-processing, in-training, intra-processing, and post-processing interventions.Source: §6.4 · pp. 59–60

6.5 Exploring Theoretical Limits

• Future work should establish theoretical fairness guarantees rather than relying primarily on empirical assessments.Theoretical analyses could both characterize existing methods and guide training procedures for models satisfying explicit fairness criteria.Source: §6.5 · p. 60
• Researchers should characterize performance–fairness trade-offs with disaggregated analyses that reveal which groups experience performance changes, treating fair outcomes as necessary rather than impediments.Pareto-frontier analysis across hyperparameter settings is one proposed way to study this relationship.Source: §6.5 · p. 60

7. Limitations

• The technical taxonomies emphasize implementation details over downstream contexts and harms, limiting engagement with structural power hierarchies and systemic inequity.The authors characterize this alignment as technical solutionism that can produce surface-level repairs without disrupting broader systems of oppression.Source: §7 · p. 60
• The survey focuses primarily on model training and evaluation, omitting broader decisions involving problem formulation, data collection, deployment, and real-world integration.Black et al. (2023) is cited as highlighting the wider scope of decisions in bias and fairness work.Source: §7 · p. 60
• The survey is limited by its focus on English-language papers.Source: §7 · p. 61

• The survey formalizes social bias and fairness in NLP, defines distinct forms of linguistic harm, and proposes initial fairness desiderata for LLMs.Source: §8 · p. 61
• It develops taxonomies for evaluation metrics, evaluation datasets, and mitigation techniques, clarifying metric–dataset relationships, dataset structures, and intervention stages.Source: §8 · p. 61
• The survey consolidates publicly available bias-evaluation datasets to improve accessibility and categorizes mitigation trends within each intervention stage.Source: §8 · p. 61
• It concludes by outlining actionable open problems and challenges intended to guide future research on measuring and reducing bias perpetuated by LLMs.Source: §8 · p. 61