CS Paper Radar

Lost in Translation: Why Lower Bias Scores Don't Mean Fairer Models

Harsh Kumar

cs.CL cs.LG ·submitted 2026-10-14 16:52 UTC · v1
code/datahot direction

Abstract

Bias audits of multilingual language models usually compare a model's average stereotype score in English with its average in another language and read the difference as a difference in fairness. We ask how much of that difference is noise. Auditing ten open models in eight languages, we measure how far scores move when English test items are merely reworded, test every language shift item by item, and check that the test catches shifts we inject on purpose. In all 70 model-language pairs the stereotype score falls below its English level, and for Hindi, Korean and Swahili the drop clearly exceeds what rewording English produces; most of it comes from StereoSet items. Judging every translation, we find that imperfect ones drop more, but faithful translations alone drop nearly as much. Still, a lower average does not show that a model is fairer. An item's English score predicts its translated score far less well than its reworded score (correlation 0.23, 0.27 on faithful translations, vs. 0.48), while scores stay as far from neutral, models agree on translated items, and within Hindi or Spanish a translated item's score correlates with its own rewording at 0.84 and 0.82. The average cannot distinguish lower bias from a different pattern of item-level scores. We also document three common pipeline errors that silently erase bias signal, and release our code.

Comment
20 pages, 6 figures. Code: https://github.com/HarshKumar3014/MultiLLMs

Closest papers

Nearest neighbours by meaning across the whole index.

Gold vs.~Translation: Are We Measuring Language, Knowledge, or Locality in Translated Benchmarks?

cs.CL ·2026-10-28 ·Klaudia-Doris Thellmann, Bernhard Stadler, Dan Saattrup Nielsen, Michael Färber, Jens Lehmann
top venuebenchmark similarity 0.66
Abstract · Published in EMNLP 2026

Translated versions of English-origin benchmarks such as MMLU are widely used to evaluate large language models in non-English languages, yet it remains unclear whether they measure the same constructs as natively created benchmarks. We frame this as a proxy validity question and compare DeepL-translated MMLU proxies against six native benchmarks in seven pairings across five Nordic and Baltic languages, evaluating 29 models on each pairing. Beyond ranking agreement, we test incremental proxy validity: does the translated score add predictive value for native performance after controlling for model size and language proficiency? Proxy validity varies substantially across benchmark pairings. Applying the same robustness criteria to all seven pairings, we find four robust effects, two marginal effects, and one null result. The strongest evidence occurs for curriculum-overlap benchmarks such as Skolprov ($ΔR^2$=.128, 95% CI [.038,.253]), whereas translated MMLU provides no significant incremental value for Sweden-specific factual knowledge in SwedishFacts (p=.242). Because SwedishFacts is also the only open-ended benchmark, we cannot fully separate the effects of local content and response format. Overall, the results indicate a clear pattern: translated benchmarks provide a better proxy when their source content overlaps with the native evaluation target, while locally anchored knowledge is less reliably captured.

0.34

From Language to Country to Valence: A Multilingual Audit of Stereotype Bias Across LLMs

cs.CL ·2026-11-02 ·Zihao Sun
similarity 0.64
Abstract · 10 pages. Author preprint; conference review ongoing

Large language models (LLMs) are increasingly considered for education, recruitment, workplace decision support, and other consequential applications, making pre-deployment social-bias detection essential. Yet major benchmarks are English-centred or translated, focus mainly on negative stereotypes, and may miss linguistic and cultural variation. We investigate gender, racial, and religious bias using 12,000 locally written Reddit posts in English, Spanish, French, and German. Only search terms are translated, preserving local context. The dataset is balanced by language and category; English gender is additionally balanced by stereotype valence. Three open-weight and three proprietary API-based models are evaluated with a two-blank design that reduces grammatical-gender cues. Religion has the highest absolute choice imbalance in every model-language condition. Spanish is highest for gender, whereas German is highest for race and religion. The evaluated API models score below the open-weight models throughout. Negative English-gender items exceed positive items; Australian and Mexican contexts exceed their Canadian and Spanish counterparts. These descriptive differences measure choice concentration, not stereotype endorsement or causal country effects. The crossed patterns show how aggregate scores can conceal category- and context-specific behaviour, motivating closer inspection before deployment.

0.15

Calibration Language Shapes Block-Influence Pruning in Multilingual Language Models

cs.CL cs.LG ·2026-10-15 ·Vedant Pandhare, Piyush Dhoka
code/data similarity 0.64
Abstract · 16 pages, 3 figures. Code and results: https://github.com/VedantPandhare/Language-Balanced-Depth-Pruning-of-Small-Multilingual-LLMs.git

Block Influence (BI) ranks Transformer blocks for depth pruning from their input-output cosine change, but the ranking depends on the calibration distribution. We study whether English calibration transfers to Indian languages in Qwen2.5-1.5B-Instruct, Llama-3.2-1B, and Gemma-2-2B. English, Hindi, Marathi, and Tamil use disjoint Wikipedia articles for calibration and evaluation; every selector receives the same 150-chunk budget, and Tamil is excluded from multilingual aggregation to test cross-script transfer. Across all three models, late-block BI is 20.7-55.5% higher for the three Indian languages than for English, and every predeclared late-minus-middle contrast has a 95% bootstrap interval above zero. This mismatch changes the low-BI blocks selected for removal despite high global rank correlations. Averaging BI over equal English, Hindi, and Marathi budgets mitigates the resulting perplexity damage for Qwen at six removed blocks and for Gemma at all tested depths, including held-out Tamil. The mitigation is not universal: on Llama, balancing improves Hindi and Marathi at two removed blocks but severely harms Tamil, while larger pruning budgets select the same blocks and collapse perplexity regardless of calibration. Rank-mean and minimax-rank aggregators do not dominate the raw mean. These results support a diagnostic claim rather than a universal pruning rule: English-only BI systematically misses late-layer activity for Indian-language text, but multilingual calibration must still be validated per architecture and pruning budget.

0.47

Benchmarking Gender Bias in Machine Translation Evaluation Metrics across Occupations

cs.CL ·2026-09-18 ·Orfeas Menis Mastromichalakis, Giorgos Filandrianos, Wafaa Mohammed, Giuseppe Attanasio, Chrysoula Zerva
published similarity 0.63
Abstract · Accepted for publication at the 11th Conference of Machine Translation (WMT26), co-located with EMNLP 2026

Gender bias remains a persistent concern in machine translation (MT), affecting both generated translations and their automatic evaluation. When a source text leaves a person's gender unspecified, translations may realize that person using masculine or feminine forms, and both MT systems and evaluation metrics may exhibit systematic preferences between these alternatives despite the source providing no basis for such a distinction. We study this behavior in the WMT 2026 Automated Translation Quality Evaluation Systems Shared Task using an occupation-balanced subset of GAMBIT+. We consider seven English-source language pairs, six from the original dataset, targeting Arabic, Czech, Greek, Icelandic, Russian, and Ukrainian, and extend the original resource with German. The subset contains 1,308 masculine/feminine translation pairs per target language, with three examples for each of the 436 ISCO-08 occupational groups. We evaluate shared-task submissions and baselines for score prediction and error annotation, examining the direction, magnitude, and frequency of gender-related differences. We find an overall tendency for masculine translations to receive higher scores, as well as differences per occupation following stereotypical gender representations, although the strength and consistency of this preference vary considerably across evaluators and languages. Our results show that gender bias remains present in MT evaluation, but that capturing its extent requires looking beyond a single aggregate measure to complementary dimensions of evaluator behavior.

0.35

Interpretable Cross-Lingual Alignment in Small Language Models: Probing Cultural and Pragmatic Reasoning in Japanese-English Bilingual LLMs

cs.CL cs.LG ·2026-08-14 ·Florian Braun
benchmark similarity 0.63
Abstract · 15 pages, no figures. Introduces the J-PragEval-v0 minimal-pair benchmark

Large language models work well on English and behave in poorly understood ways on languages typologically far from it. Japanese is a clean example, where evaluation still leans on translation quality and JGLUE-style benchmarks, which roll lexical, syntactic and pragmatic competence into a single score. The phenomena on which general-purpose models fail Japanese users are pragmatic: honorifics, in-group and out-group reference, context-sensitive politeness, zero anaphora. I introduce J-PragEval-v0, a minimal-pair benchmark isolating four such phenomena from surface fluency, and combine it with linear probes and teacher-forced log-probability evaluation to ask where inside TinySwallow-1.5B (28 layers, hidden size 1536) the corresponding contrasts live. The four features split three ways. Honorific register sits cleanly in the residual stream: 0.96 balanced accuracy at layer 15, and the model flips its preferred continuation with the scenario on 93 percent of items. Implicit subject and in-group reference are not linearly decodable at the final prompt token (0.48 and 0.38), yet flip rates are 0.77 and 0.79, so the contrast is worked out during generation rather than stored at the prompt. Indirect refusal is the negative case: 0.95 probe accuracy collapsing to a 0.43 flip rate under length-normalised teacher forcing, because the current minimal pairs conflate politeness with continuation length. I also specify Pragmatic Representation Steering, a parameter-free inference-time method that edits residual-stream activations along the class-mean-difference directions probing identifies. Feasibility is argued indirectly rather than demonstrated: the contrastive activation addition baseline, the same geometry the method would inject, recovers probe accuracy within one to two points of logistic regression wherever a linear signal exists. Scaling to Llama-3.1-Swallow-8B is the next step.

0.40

Disentangling Question, Response, and Verifier Languages in Multilingual P(True) Self-Evaluation

cs.CL ·2026-10-23 ·Jewon Yeom, Jaewon Sok, Seonghyeon Park, Jeongjae Park, Hwiyeong Lee +1 more
benchmarkhot direction similarity 0.63
Abstract

When a language model is asked whether its own answer is correct, three languages are involved: that of the question, that requested for the answer, and that of the verifier instruction asking for the judgment. Multilingual confidence evaluations seldom separate the three. We vary them independently in a six-language factorial design with ten open-weight models (3B-14B) on LSQA, MMMLU and TriviaQA, holding each answer fixed across verifier languages and measuring confidence with a first-token P(True) probe. The three languages move different parts of the evaluation: which of them governs accuracy depends on the benchmark, matching the response language to the question or verifier language sharpens discrimination but not calibration, and the verifier language alone shifts calibration error by up to 0.04 without changing accuracy. The scorer matters as well: an unvalidated judge would have shrunk the MMMLU English-Hindi accuracy gap from 0.198 to 0.038. Multilingual confidence results, model rankings included, are therefore comparable only when all three languages are reported and the scorer is validated.

0.29

WinoQueer-NL: Assessing Bias in Dutch Language Models toward LGBTQ+ Identities

cs.CL ·2026-09-02 ·Jiska Beuk, Gerasimos Spanakis
benchmarkcode/data similarity 0.61
Abstract · accepted at 7th Workshop on Gender Bias in Natural Language Processing (GeBNLP2026) @ AACL2026. Dataset available via https://github.com/jer…

While English language models have been widely examined for anti-queer bias, Dutch models remain understudied. To address this gap, we developed a culturally and linguistically adapted Dutch dataset based on the English WinoQueer benchmark, containing pairs of stereotypical and counter-stereotypical sentences. To validate and expand it, we conducted an online survey with 43 Dutch queer participants, confirming 145 of 171 stereotypes as culturally relevant and identifying 22 new biases through free-text responses. The final released dataset, comprising 42,906 sentences, was evaluated using a range of Dutch-specific and multilingual models, including both masked language models (MLMs) and autoregressive language models (ARLMs), with bias measured via a score comparing log-likelihoods of stereotypical versus counter-stereotypical sentences. While the mean bias score across models appeared neutral (~50%), closer analysis revealed significant disparities: some models favored stereotypical sentences up to 97% of the time for transgender identities, but only 6% of the time for gay-related pairs, with transgender and non-binary identities consistently receiving the highest bias scores. Our findings highlight the importance of culturally grounded datasets for evaluating and mitigating biases that disproportionately impact marginalized groups in Dutch language models.

0.40

Fair Fact-Checking: Closing the Cross-Lingual Gap in LLM Factual Judgement with RoSh

cs.CL cs.AI ·2026-09-28 ·Muhammad Ahmad, Fatemeh Seyedin, Adrian Weller, Dongwon Lee, Mahmoudreza Babaei
similarity 0.60
Abstract · 23 pages, 3 figures

Misinformation on social media remains a critical problem, and more and more people settle it by asking a language model instead of a fact checker. Whether models judge such claims reliably is debated; whether they judge them equally well in every language people ask in has gone almost unasked. We test eight models from five families, 3B to 70B, on 1,500 encyclopedic factual claims that exist in identical form in eight languages. English is judged better than every other language on every model, and the gap is widest on the smallest ones, where Llama-3B on Arabic is no better than guessing. Existing remedies retrain on more multilingual data or fit an unconstrained map between language representations, and neither asks whether the model already holds the answer and simply fails to say it. It largely does: a linear probe recovers the truth from the very activations the model fails to express. We propose RoSh, a per-language shift and rotation of the residual stream, computed in closed form at three layers, with no training and no weight modified. It improves every model and closes 75% of the gap on average, helping most where the model was worst: Arabic on Llama-3B goes from chance to nearly the English level, and a fifth fewer of the claims answered correctly in English are lost in translation. What remains is no longer a read-out failure: afterwards the head recovers as much of what is encoded outside English as it does in English. An unconstrained map fitted on the same pairs falls below the untouched baseline, so the orthogonality constraint is doing the work, and every model clears a scrambled-correspondence control and ten further controls. On the two benchmarks of the closest inference-time method, latent-space intervention, run with its own data and metric code, RoSh's gains are five to thirteen times larger.

0.12