cs.CLAug 24, 2026

The Dialect Tax: Dialectal Biases Persist throughout the Language Modeling Pipeline

Authors: Elle

Organizations: University of Oxford, Department of Computer Science

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Systematic dialectal performance gaps in language models (LMs) are well documented, but the source of these disparities within the modern language modeling pipeline remains unclear. Our study traces this "dialect tax" across the natural language processing pipeline. Using parallel English dialect corpora that hold meaning fixed while varying surface form, we first confirm that LMs recognize matched Standard American English (SAE) and dialectal texts as semantically equivalent. However, we discover further representational gaps corresponding to downstream performance gaps. Across model families and generations, modern LMs still encode dialectal texts unequally during tokenization, pre-training, post-training, and inference. Strikingly, bypassing traditional subword segmentation via a character-level counterfactual tokenizer removes neither input and output asymmetries nor dialectal accuracy gaps. During pre-training, dialect pairs induce more divergent gradient updates than pairs of entirely unrelated SAE documents, indicating that models find semantically equivalent dialectal content harder to learn from than unrelated SAE documents. During post-training, reward models show contextual, unstable dialect preferences, assigning higher values to isolated AAVE-exclusive tokens than to SAE-exclusive tokens, while full reasoning contexts receive task- and model-dependent dialect penalties. Overall, our findings suggest that the dialect tax is encoded and accumulated not by any one step in isolation, but at every step of the language modeling process.

Figures & tables

Figure 1: Models understand semantic equivalence yet penalize surface form. We visualize the semantic equivalence of various text transformations on MultiVALUE . All dialect pairs achieve high similarities exceeding every perturbation and translation baseline.
Figure 2: Modern tokenizers consistently exhibit dialectal biases. ( 2(a) ) The mean fertility of the three tokenization algorithms on ParallelAAVE shows a statistically significant difference between SAE and AAVE dialects. ( 2(b) ) The ratio of BPE token lengths on various dialects to that of SAE on MultiVALUE reveals a consistent dialectal tokenization performance gap that loosely parallels current income gaps of minority groups within the US (§ D.1 ).
Figure 3: Character tokenization does not systematically narrow dialect gaps. We plot per-model SAE −- AAVE effect sizes under canonical (blue) and character-level (orange) tokenization, split by reasoning strategy ( △=CoT\triangle=\text{CoT} , ◯=nai¨ve\bigcirc=\text{na\"{i}ve} ). Grey lines connect paired points for each model. Left: The accuracy gap does not systematically change under character tokenization. Center: The input entropy gap ΔH\Delta_{H} changes inconsistently, confirming that character tokenization fails to equalize how the model processes dialect inputs. Right: The output entropy gap ΔH\Delta_{H} persists under both tokenizations, suggesting generation-time dialect bias is independent of the tokenizer.
 LlamaGemmaQwen
 1B3B8B1B4B12B1.7B4B8B
Canonical (%)99.099.999.499.899.8100.099.9100.0100.0
Character (%)94.598.399.686.499.599.598.597.397.5
Δ\Delta−4.5-4.5−1.6-1.6+0.2+0.2−13.4-13.4−0.3-0.3−0.5-0.5−1.3-1.3−2.7-2.7−2.5-2.5
Table 1: Dialects are linearly decodable from hidden states under both tokenizations. We list the five-fold cross-validation accuracy of logistic regression predicting dialect from the answer-step hidden state. Character tokenization slightly reduces separability, but all models remain far above chance (50%).
Figure 4: Dialectal form outweighs semantic mismatch in gradient geometry. Violin plots of the gradient cosine similarity between paired and unrelated documents on ReDial reveal that matched SAE-AAVE pairs produce lower similarity than unrelated SAE-SAE pairs, implying that a meaning-preserving dialect shift can perturb the training signal more than changing the document content within SAE. In the plot, an LM name is indicated by its first-letter abbreviation and size.
Figure 5: Dialectal text incurs higher prediction loss. We plot the mean cross-entropy loss by dialect. All nine models assign significantly higher loss to AAVE inputs than to their SAE counterparts (one-sided paired Wilcoxon test, n=1,200n=1{,}200 per model, Bonferroni-corrected p<0.001p<0.001 ), with per-model mean gaps of 0.470.47 to 0.660.66 nats (Cohen’s dz∈[2.12,2.55]d_{z}\in[2.12,2.55] ). An LM name is indicated by its first-letter abbreviation and size.
Paired condition with SAEμ+(↑)\mu^{+}(\uparrow)zz
Baseline (unrelated SAE)0.570.570.000.00
Capitalize (alternating)0.290.29−3.36-3.36
Capitalize (random)0.340.34−2.88-2.88
Drop ( P=0.15\mathbb{P}=0.15 )0.360.36−2.58-2.58
Insert ( P=0.05\mathbb{P}=0.05 )0.470.47−1.40-1.40
Swap ( P=0.05\mathbb{P}=0.05 )0.500.50−1.06-1.06
Table 2: Dialect gradient divergence sits within the range of character-level perturbations. Across nine base models and four ReDial tasks, we report the grand-mean paired cosine similarity ( μ+\mu^{+} ) and zz -score against the unrelated-SAE baseline. Higher μ+\mu^{+} indicates greater similarity to SAE gradients; lower values indicate stronger gradient divergence.
Figure 6: Transformations result in similar hidden-state similarity curves. Layer-wise cosine similarities between SAE hidden states and those of transformed texts are pooled across MultiVALUE and ParallelAAVE . All text transformations follow similar trajectories across normalized LM layers, with dialectal transformations most similar to SAE.
Figure 7: Hidden-state similarity \mathchoice{\mathrel{\hbox to0.0pt{\kern 5.0pt\kern-5.27776pt\displaystyle\not\hss}{\Rightarrow}}}{\mathrel{\hbox to0.0pt{\kern 5.0pt\kern-5.27776pt\textstyle\not\hss}{\Rightarrow}}}{\mathrel{\hbox to0.0pt{\kern 3.98611pt\kern-4.45831pt\scriptstyle\not\hss}{\Rightarrow}}}{\mathrel{\hbox to0.0pt{\kern 3.40282pt\kern-3.95834pt\scriptscriptstyle\not\hss}{\Rightarrow}}} downstream accuracy. On the ReDial dataset, we plot each transformation’s final-layer hidden-state cosine similarity to SAE against its change in answer accuracy. Among semantic-meaning-preserving transformations, SAE-to-AAVE shows the highest similarity to the original SAE representation but the lowest downstream accuracy.
Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

CategorySourceItems
Algorithm (26%)HumanEval ( Chen et al., 2021 )164
MBPP (Sanitized) ( Austin et al., 2021 )149
Logic (30%)FOLIO ( Han et al., 2024 ; Wu et al., 2024 )162
LogicBench ( Parmar et al., 2024 )200
Math (25%)GSM8K ( Cobbe et al., 2021 )150
SVAMP ( Patel et al., 2021 )150
Appendix
Table 3: Source datasets used to reconstruct ReDial .
DatasetSourcePairsDialects
ParallelAAVEGroenwold et al. (2020)20192019SAE, AAVE
MultiVALUEZiems* et al. (2023)429429SAE, AAVE, Appalachian, Chicano, Indian, Singapore
Appendix
Table 4: Datasets for parallel SAE and AAVE texts.
TokenizationNameModel / Encoding
BPEGPT-5o200k_base (tiktoken)
GPT-2openai-community/gpt2
Gemmagoogle/gemma-3-27b-it
Llamameta-llama/Llama-3.3-70B-Instruct
QwenQwen/Qwen3-32B
UnigramT5t5-small
Appendix
Table 5: Tokenizer names used in our analysis.
FamilyNameSizeTypeModel ID
LlamaLlama 3.21BBasemeta-llama/Llama-3.2-1B
Llama-3.21BInstructmeta-llama/Llama-3.2-1B-Instruct
Llama-3.23BBasemeta-llama/Llama-3.2-3B
Llama-3.23BInstructmeta-llama/Llama-3.2-3B-Instruct
Llama-3.18BBasemeta-llama/Llama-3.1-8B
Llama-3.18BInstructmeta-llama/Llama-3.1-8B-Instruct
Appendix
Table 6: Language models used in our analysis.
ProviderBase ModelSizeModel ID
SkyworkLlama-3.2 (Instruct)3BSkywork/Skywork-Reward-V2-Llama-3.2-3B
Llama-3.1 (Instruct)8BSkywork/Skywork-Reward-V2-Llama-3.1-8B
Qwen-3 (Instruct)4BSkywork/Skywork-Reward-V2-Qwen3-4B
Qwen-3 (Instruct)8BSkywork/Skywork-Reward-V2-Qwen3-8B
Gemma-2 (Instruct)27BSkywork/Skywork-Reward-Gemma-2-27B
QRMLlama-3.1 (Instruct)8Bnicolinho/QRM-Llama3.1-8B-v2
Appendix
Table 7: Reward models used in our analysis.
Resource LevelLanguageGoogle Translate Code
HighChinesezh-CN
Frenchfr
MidHindihi
Polishpl
LowKhmerkm
Yorubayo
Appendix
Table 8: Translation transformation details.
 MultiVALUEParallelAAVEReDial
ConditionRatiofrac >> 1Ratiofrac >> 1Ratiofrac >> 1
Drop ( P=0.15\mathbb{P}=0.15 )12.5312.531.001.009.649.641.001.002.062.061.001.00
Singapore4.924.921.001.00––––
Insert ( P=0.05\mathbb{P}=0.05 )4.414.411.001.003.743.740.980.981.581.581.001.00
Indian3.993.991.001.00––––
Drop ( P=0.05\mathbb{P}=0.05 )3.523.521.001.003.063.060.980.981.431.431.001.00
Appendix
Table 9: Paired per-token perplexity ratio vs SAE. We compare the perplexity ratios for character perturbations, translations, and dialects ( italicized ). Each row pairs the transformed text to its matched SAE text by unique_id within a model (ratio =exp⁡(CEcond−CEsae)=\exp(\text{CE}_{\text{cond}}-\text{CE}_{\text{sae}}) ), aggregated as the median of per-model medians across the models listed in Table 6 . The mean fraction of paired items the model finds harder than SAE is indicated under “frac >> 1” , where a ratio above 11 means the model is more surprised by the surface form than by SAE despite preserved meaning.
Figure 8: Models understand semantic equivalence yet penalize surface form. We find evidence of semantic invariance under surface-form transformations, as shown on the MultiVALUE corpus. ( 8(a) ) We visualize the semantic equivalence of various text transformations. All dialect pairs achieve high cosine similarities above 0.970.97 , which exceed every perturbation and translation baseline. ( 8(b) ) We plot semantic equivalence (cosine similarity at d=768d=768 ) against the tokenization tax ( Δfertility=E[fertilitydialect]−E[fertilitySAE]\Delta\text{fertility}=\mathbb{E}[\text{fertility}_{\text{dialect}}]-\mathbb{E}[\text{fertility}_{\text{SAE}}] ) for each dialect-SAE pair. Samples in the upper-right exhibit high semantic equivalence but higher tokenization cost, where meaning is preserved while a tax is imposed.
Figure 9: Models understand semantic equivalence yet penalize surface form. We find evidence of semantic invariance under surface-form transformations, as shown on the ParallelAAVE corpus. ( 9(a) ) We visualize the semantic equivalence of various text transformations. The AAVE pairs achieve high cosine similarities, which exceed most perturbation and translation baselines. ( 9(b) ) We plot semantic equivalence (cosine similarity at d=768d=768 ) against the tokenization tax ( Δfertility=E[fertilityAAVE]−E[fertilitySAE]\Delta\text{fertility}=\mathbb{E}[\text{fertility}_{\text{AAVE}}]-\mathbb{E}[\text{fertility}_{\text{SAE}}] ) for each AAVE-SAE pair. Samples in the upper-right exhibit high semantic equivalence but higher tokenization cost, where meaning is preserved while a tax is imposed.
MetricWhat does it measure?
Average tokens per wordAverage number of tokens corresponding with a single real word without punctuations
Average types per wordAverage number of types corresponding with a single real word without punctuations
Character lengthNumber of characters in the string
Fertility ( Ács, 2019 )Average number of tokens corresponding with a single real word
P(in vocabulary)Proportion of words in the tokenizer vocabulary
Token lengthNumber of tokens in the string
Appendix
Table 10: Tokenization metrics used to measure bias.
Figure 10: Ratio of dialect to SAE token lengths on MultiVALUE . Dotted lines indicate token-length parity with the paired SAE text. We see similar tokenization bias ranking on every measured metric for the six dialects in the dataset (Appalachian >> AAVE >> Chicano >> SAE >> Indian >> Singapore). While some dialects (AAVE, Appalachian, Chicano) have increased token lengths compared to SAE, other dialects (Indian, Singapore) have decreased token lengths compared to SAE. We find that the same dialectal token bias remains present across all three tokenization strategies.
Figure 11: Dialectal token bias and income. We demarcate token parity with y=1y=1 and the US median household income with x=\80,610$ . Compared to Standard American English (SAE) speakers, speakers of African American Vernacular English (AAVE), Appalachian, and Chicano dialects who earn less than the median income achieve less efficient tokenization. Meanwhile, speakers of Indian and Singaporean dialects who earn more than the median income achieve more efficient tokenization than SAE speakers.
DialectIncomeSource
AAVE\56,490$Guzman (2024)
Appalachian\61,688$Srygley et al. (2024)
Chicano\65,540$Guzman (2024)
Indian\166,200$Lee and Batalova (2026)
SAE\80,610$Guzman (2024)
Singaporean\134,818$ZipAtlas (2025)
Appendix
Table 11: US median household income (USD).
Figure 12: Tokenizer metric timeline. We plot a subset of the token metrics in Table 10 over time.
FamilySizerr
Llama-3 Base1B−0.119∗∗∗-0.119^{***}
3B−0.053-0.053
8B−0.091∗∗-0.091^{**}
Gemma-3 Base1B−0.108∗∗∗-0.108^{***}
4B−0.105∗∗∗-0.105^{***}
12B−0.066-0.066
Appendix
Table 13: Point-biserial correlation between paired SAE-AAVE gradient similarity si+s_{i}^{+} and a binary indicator for both-dialect correctness, by model. Per-model sample size is n=1,200n=1{,}200 , and pooled sample size is n=10,800n=10{,}800 . We denote significance by ∗∗∗ p<0.001{}^{***}\,p<0.001 and ∗∗ p<0.01{}^{**}\,p<0.01 . Pooled rr is negligible, but the per-model breakdown reveals heterogeneity.
Reward modelΔr‾\overline{\Delta r}tt
Skywork Llama 3B−0.03-0.03−0.67-0.67
Skywork Qwen 4B+0.32+0.32+7.54∗∗∗+7.54^{***}
Skywork Llama 8B−0.22-0.22−3.27∗∗-3.27^{**}
Skywork Qwen 8B+0.48+0.48+14.42∗∗∗+14.42^{***}
Skywork Gemma 27B−0.03-0.03−0.31-0.31
QRM Llama 8B+0.01+0.01+4.72∗∗∗+4.72^{***}
Appendix
Table 14: Per-RM dialect gap. Δr‾=E[r(xSAE,y)−r(xAAVE,y)]\overline{\Delta r}=\mathbb{E}[r(x^{\text{SAE}},y)-r(x^{\text{AAVE}},y)] over ReDial sample-level pairs ( n=1,200n=1{,}200 per RM, i.e. 300 per task ×\times 4 tasks), and one-sample tt -statistic against E[Δr]=0\mathbb{E}[\Delta r]=0 . Significance is denoted by ∗∗∗ p<0.001{}^{***}\,p<0.001 , ∗∗ p<0.01{}^{**}\,p<0.01 , ∗ p<0.05{}^{*}\,p<0.05 .
TaskΔr‾\overline{\Delta r}tt
Algorithm+0.43+0.43+13.73∗∗∗+13.73^{***}
Math+0.04+0.04+1.18+1.18
Logic−0.08-0.08−2.92∗∗-2.92^{**}
Planning−0.14-0.14−4.92∗∗∗-4.92^{***}
Appendix
Table 15: Per-task dialect gap, pooled across RMs. One-sample tt -test of E[Δr]=0\mathbb{E}[\Delta r]=0 within task. Significance is denoted by ∗∗∗ p<0.001{}^{***}\,p<0.001 , ∗∗ p<0.01{}^{**}\,p<0.01 .
CorpusDialectsnSAEn_{\text{SAE}}ndialectn_{\text{dialect}}r‾SAE\overline{r}_{\text{SAE}}r‾dialect\overline{r}_{\text{dialect}}r‾SAE−r‾dialect\overline{r}_{\text{SAE}}-\overline{r}_{\text{dialect}}d‾RM\overline{d}_{\text{RM}}pp
ReDial{AAVE}121,780163,120−4.47-4.47−4.10-4.10−0.37∗∗∗-0.37^{***}−0.17-0.179×10−319\times 10^{-31}
ParallelAAVE{AAVE}91,380163,420−4.42-4.42−3.83-3.83−0.58∗∗∗-0.58^{***}−0.29-0.291×10−671\times 10^{-67}
MultiVALUE{AAVE, Appal., Chic., Indian, Sing.}141,930132,130−5.17-5.17−4.74-4.74−0.43∗∗∗-0.43^{***}−0.24-0.249×10−399\times 10^{-39}
Pooled (all) 355,090458,670−4.74-4.74−4.19-4.19−0.55∗∗∗-0.55^{***}−0.27-0.277×10−1857\times 10^{-185}
Appendix
Table 16: Per-corpus token-level dialect gap. For each corpus, we identify subword tokens that appear exclusively in tokenized SAE vs. dialect text within each (RM, tokenizer) pairing, and score them under a fixed prompt. r‾SAE\overline{r}_{\text{SAE}} and r‾dial\overline{r}_{\text{dial}} are the mean reward scores over the dialect-exclusive vocabularies. r‾SAE−r‾dial\overline{r}_{\text{SAE}}-\overline{r}_{\text{dial}} is the raw gap. d‾RM\overline{d}_{\text{RM}} is Cohen’s dd computed within each RM (using that RM’s pooled score standard deviation) and then averaged across the ten RMs. This normalizes for the substantial cross-RM scale differences (per-RM σ∈[0.06,3.97]\sigma\in[0.06,3.97] ). The bottom row pools all three corpora together. Independent two-sample tt -tests, with significance denoted by ∗∗∗ p<0.001{}^{***}\,p<0.001 .
Reward modelAAVEAppalachianChicanoIndianSingaporeanPooled
Skywork Llama 3B−0.45∗∗∗-0.45^{***} ( −0.25-0.25 )+0.11∗+0.11^{*} ( +0.06+0.06 )−0.71∗∗∗-0.71^{***} ( −0.43-0.43 )−0.06-0.06 ( −0.03-0.03 )−0.88∗∗∗-0.88^{***} ( −0.49-0.49 )−0.51∗∗∗-0.51^{***} ( −0.28-0.28 )
Skywork Qwen 4B−0.69∗∗∗-0.69^{***} ( −0.31-0.31 )−0.32∗∗∗-0.32^{***} ( −0.15-0.15 )−0.71∗∗∗-0.71^{***} ( −0.37-0.37 )−0.13∗-0.13^{*} ( −0.06-0.06 )−1.10∗∗∗-1.10^{***} ( −0.47-0.47 )−0.78∗∗∗-0.78^{***} ( −0.34-0.34 )
Skywork Llama 8B−1.26∗∗∗-1.26^{***} ( −0.38-0.38 )−0.09-0.09 ( −0.03-0.03 )−1.49∗∗∗-1.49^{***} ( −0.49-0.49 )−0.28∗∗-0.28^{**} ( −0.08-0.08 )−1.34∗∗∗-1.34^{***} ( −0.38-0.38 )−1.26∗∗∗-1.26^{***} ( −0.37-0.37 )
Skywork Qwen 8B−0.80∗∗∗-0.80^{***} ( −0.40-0.40 )−0.16∗∗-0.16^{**} ( −0.09-0.09 )−0.67∗∗∗-0.67^{***} ( −0.35-0.35 )+0.00+0.00 ( +0.00+0.00 )−0.61∗∗∗-0.61^{***} ( −0.29-0.29 )−0.73∗∗∗-0.73^{***} ( −0.36-0.36 )
Skywork Gemma 27B−1.04∗∗∗-1.04^{***} ( −0.27-0.27 )−1.74∗∗∗-1.74^{***} ( −0.45-0.45 )−1.97∗∗∗-1.97^{***} ( −0.54-0.54 )+2.15∗∗∗\mathbf{+2.15^{***}} (+0.57)\mathbf{(+0.57)}−2.08∗∗∗-2.08^{***} ( −0.52-0.52 )−1.17∗∗∗-1.17^{***} ( −0.30-0.30 )
QRM Llama 8B+0.00∗∗∗+0.00^{***} ( +0.04+0.04 )−0.00-0.00 ( −0.01-0.01 )−0.02∗∗∗-0.02^{***} ( −0.37-0.37 )−0.00-0.00 ( −0.05-0.05 )−0.01∗∗∗-0.01^{***} ( −0.19-0.19 )−0.00∗-0.00^{*} ( −0.02-0.02 )
Appendix
Table 17: Per-RM per-dialect token-level dialect gap. Each cell shows the raw reward gap r‾SAE−r‾dial\overline{r}_{\text{SAE}}-\overline{r}_{\text{dial}} with Cohen’s dd in parentheses. Negative values indicate the RM scores dialect-exclusive tokens higher than SAE-exclusive ones. AAVE-exclusive tokens come from ReDial , ParallelAAVE , and MultiVALUE , and other dialect-exclusive tokens come from MultiVALUE . Cohen’s dd uses each RM’s pooled score standard deviation as denominator, so raw gaps are not directly comparable across RMs while dd values are. The bolded cell is the only one flipping the dialect-favoring pattern. The “Pooled” column aggregates a given RM across all five dialects, and the “Mean across RMs” row is the unweighted average across the ten RMs. We run independent two-sample tt -tests and denote the significance: ∗∗∗ p<0.001{}^{***}\,p<0.001 , ∗∗ p<0.01{}^{**}\,p<0.01 , ∗ p<0.05{}^{*}\,p<0.05 .
   Raw scoresWithin-RM standardized
DialectnSAEn_{\text{SAE}}ndialn_{\text{dial}}r‾SAE−r‾dial\overline{r}_{\text{SAE}}-\overline{r}_{\text{dial}}ppddpp
AAVE239,000351,140−0.50-0.506×10−1156\times 10^{-115}−0.24-0.24<10−300<10^{-300}
Appalachian24,98031,540−0.29-0.298×10−58\times 10^{-5}−0.13-0.135×10−645\times 10^{-64}
Chicano8,9606,640−0.73-0.733×10−73\times 10^{-7}−0.39-0.392×10−1652\times 10^{-165}
Indian27,20032,710+0.11+0.110.140.14−0.03-0.034×10−44\times 10^{-4}
Singaporean54,95036,640−0.79-0.796×10−406\times 10^{-40}−0.40-0.40<10−300<10^{-300}
Appendix
Table 18: Per-dialect token-level gap, before and after within-RM standardization. Raw scores: pooled two-sample tt -test on raw reward scores. Note that per-RM output scales differ, so the test is heavily influenced by the high- σ\sigma Skywork models. Within-RM standardized: each RM’s scores are zz -scored using its own mean and standard deviation, then pooled across RMs before the tt -test. The dd column reports the gap in standard deviations, which is mathematically equivalent to the unweighted mean Cohen’s dd across the ten RMs (within-RM effect size).
Figure 13: Hidden-state similarity drops at the final layers for every model on MultiVALUE . Per-model layer-wise cosine similarity between SAE and dialect hidden states on MultiVALUE . Grey dashed curves show each of the five dialects (AAVE, Appalachian, Chicano, Indian, Singaporean) separately; colored curves are the per-model mean across the five dialects (solid = base, dashed = instruct), with color encoding family.
Figure 14: Final-layer hidden-state similarity. The final-layer hidden-state similarity to that of SAE is pooled across MultiVALUE and ParallelAAVE .
More in this topic

Dialect Bias in Language Models

See all 14 papers

Explore similar work

CardsList

  1. Side-by-side Comparison Amplifies Dialect Bias in Language Models

    May 23, 2026Kritee Kondapally, Claire J. Smerdon, Pooja C. Patel +5Social Bias in Language ModelsAlgorithmic Fairness

    University of Southern California, USA

  2. Some Dialects Are More Equal Than Others: Non-Prestigious Arabic Dialectal Bias in LLMs

    Sep 21, 2026Mai Mohamed Eida, Ryan Dolan, Paul de Nijs +1Arabic NLPSocial Bias in Language Models

    University of Illinois Urbana Champaign

  3. DiaLLM: An Investigation into the Robustness-Generation Gap in English Dialect Adaptation

    Jul 8, 2026Jordan Painter, Dipankar Srirag, Adarsh Kappiyath +3Language ModelingLLM Alignment

    Institute for People-Centred AI, University of Surrey, Surrey, United Kingdom · University of New South Wales, Sydney, Australia