Women remain underrepresented in computer science, yet little is known about how gendered associations within the field have changed over time. This study examines longitudinal patterns in gender–competence associations across academic and practitioner discourse from 2010–2025 using 210 Word2Vec models trained on 9.6 billion tokens from arXiv computer science papers as well as Stack Overflow posts and comments. Female-competence associations were measured using relative norm distance (RND) between female, male, and competence wordsets. In academic discourse, relative female-competence associations decreased significantly over time (𝜏= -0.717, p < 0.001), reflecting a shift towards stronger relative-male competence association. In contrast, practitioner discourse showed no significant longitudinal trend (𝜏 = -0.250, p = 0.192). Academic gender-competence trends were significantly associated with women’s representation in the U.S. computer science workforce whereas practitioner trends were not. These findings demonstrate that gendered linguistic associations have not evolved uniformly across computer science and highlight the importance of discourse context when evaluating longitudinal changes in gendered language.

Investigations of gender-neutral and non-binary language in Romance languages often fail to take into account learner language. There is a lack of material to study how grammatical error correction models handle non-binary language which uses new forms or morphemes. In this paper, we offer Valicə a gender-neutralized version of the Valico treebank, containing Italian learner texts and their corrections. We present the procedure to gender-neutralize Italian sentences, both grammatically correct and incorrect. We implement two different systems to identify which tokens should be neutralized: a rule-based system and a machine learning system, both of which prove equally efficient. The token swap is rule-based, using the neomorphemes -ə and -ɜ to signal singular and plural gender-neutrality. The Valicə dataset can be used to evaluate or train language learning applications for learners of Italian.

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 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.

Third-person singular pronouns have long been used to study stereotypical biases in language models and to test their abilities to reason about reference. More recently, the interplay between reasoning and bias has been investigated with the task of pronoun fidelity, which assesses models’ abilities to correctly reuse a previously-specified pronoun for a discourse entity, independent of other potentially distracting discourse entities mentioned in between. However, such research focuses on English, which is a language with limited grammatical gender and almost no gender agreement. In this paper we contribute a novel, large-scale dataset, **GRUFF**, to measure pronoun fidelity in German, covering four different gender agreement systems in nouns, and four sets of pronouns. With this dataset, we show that LLMs display strong grammatical agreement for masculine and feminine entities in the absence of explicit context, but not for neopronouns _xier_ and _en_. In the presence of distractors, however, models degrade on all pronoun sets. Finally, we show that occupational stereotypes in this context are poorly correlated across grammatical cases, and across most models, except ones with closely related architectures. We release all code and data to encourage further work on gender-inclusive language and referential reasoning in German.

Large language models are increasingly asked to weigh in on interpersonal disputes and apportion blame, a setting where gender bias in fault attribution would carry direct relational stakes. To test whether such judgments depend on the disputants’ gender identities, we introduce 1,458 factorially controlled relationship-dispute vignettes crossing 9 relationship-norm families, 2 severity levels, and a full set of male, female, nonbinary agent and partner identities. Five LLMs from five providers rate each on a single 0–7 fault scale. Male agents are rated more at fault than otherwise identical female agents (+0.16 points, matched-pair dz = 0.316), the direction prior work reports, but the effect is not uniform: it varies significantly by relationship-norm family and by model. A prespecified ambivalent-sexism account of which families should show the largest effect fails its confirmatory test. Nonbinary-tagged agents are rated much closer to female-tagged than to male-tagged agents (0.03 vs. 0.12 points), while the partner’s identity forms a complementary axis with the opposite sign. The two interact: a male partner lowers a female or nonbinary agent’s rating, with no detectable change for male-tagged agents, a pattern a binary same- versus mixed-gender comparison cannot distinguish from a pairing effect. The paired contrasts retain their direction on the 45 held-out scenarios, and every contrast significant on the full set except MM vs. FF remains significant there. Excluding Claude Sonnet 5, which assisted in drafting some vignettes, preserves significance except for the very small agent female–nonbinary contrast.

Gender bias in NLP has mostly been studied through the lens of stereotypes, i.e. associations between gender and occupations or personality traits. In this position paper, we argue that the rise of conversational large language models (LLMs) and open-ended generation stressed another important problem that remains largely unaddressed: masculine generics bias. In many languages, including German, French or Spanish, masculine forms are used to refer to mixed-gender groups or to people whose gender is unknown or non-binary (masculine generics, or MG for short). Decades of psycholinguistic research show that these forms are not interpreted as truly generic and instead lead people to have more male-centric mental representations. As LLMs are adopted by an ever-growing number of users worldwide for a wide range of tasks, this bias can now be perpetuated in everyday interactions, all the more since preliminary studies already show that LLMs lean heavily on MG and are reluctant to use gender-fair language spontaneously. We review the evidence for why it matters, explain why existing mitigation approaches are not enough for conversational LLMs, show that GFL in German, French and Spanish is lacking in LLM pre-training corpora, and outline 6 research directions towards building multilingual, linguistically gender-fair conversational LLMs.

Do language models adapt their linguistic style in an unbiased and socially appropriate way? We examine this issue in Japanese, a language with complex sociolinguistic constraints on stylistic adaptation between interlocutors, and we focus on the interplay between politeness and gendered linguistic styles. In terms of phenomena, we analyze morphologically encoded levels of formality alongside gender-indexical cues. In terms of methods, we analyze model behavior as well as human perceptions of generated outputs. Concretely, we first curate a novel set of 900 topic-balanced prompts with systematic formality and gendered-style variants. We then use these prompts to examine the behavior of four models, finding weak-to-moderate but consistent formality adaptation, with stronger adaptation to masculine-style than to feminine-style questions. Finally, we conduct a human perception study, showing that formality-mismatched responses are perceived as less appropriate than matched responses, with judgments strongly shaped by situational context. Taken together, our findings highlight the value of stylistic adaptation that is broadly invariant to gender-indexical cues and sensitive to situational context, with direct implications for conversational system design.

Implicit misogyny often requires background knowledge to be understood. To address this challenge, we propose a Retrieval-Augmented Generation (RAG) framework for detecting and explaining implicit misogyny in English and Italian, leveraging encyclopedic, slang, and feminist knowledge sources. Using the ImplicIT-Mis and SBIC+ datasets, we evaluate the contribution of each knowledge source to detection and explanation quality, investigating also whether RAG addresses known LLM failure modes in implicit misogyny. Results show that RAG improves recall on Italian data, but does not improve English detection performance. However, human evaluation shows that RAG-based explanations are preferred over no-RAG explanations in both languages. Feminist literature emerges as the most consistently useful source for detection and explanation.

Large language models (LLMs) are widely used for paraphrase generation, making it important to evaluate their ability to produce fair outputs under explicit debiasing instructions, particularly for under-resourced languages such as Bangla. Existing work has focused on paraphrase generation and detection but lacks benchmarks for evaluating instruction-guided debiasing. We introduce BaP-Bench, a bias-aware Bangla paraphrase benchmark constructed using four LLMs to evaluate instruction-guided mitigation of stance/sentiment toward women (as a proxy for gender bias). To enable large-scale evaluation, we employ automatic classifier-based annotation to assess the effectiveness of LLM-generated debiased paraphrases. We further demonstrate the utility of BaP-Bench as a downstream benchmark by evaluating transformer-based models on paraphrase detection under debiasing instructions. Results show that LLMs vary considerably in their ability to follow debiasing instructions while preserving paraphrase quality, and that transformer-based models achieve strong performance on the downstream task. BaP-Bench provides a foundation for studying instruction-guided fairness in Bangla paraphrasing and supports future research on bias-aware paraphrase generation and evaluation.

We investigate whether product ratings and source-text emotions are associated with first-person gender choices in English-to-Russian and English-to-Croatian translations by professional translators, students, and three LLMs (GPT-4o mini, Llama, and Mistral). We score love, sadness, and anger in the source reviews with a RoBERTa classifier fine-tuned on GoEmotions and fit logistic mixed-effects models of feminine versus masculine forms, accounting for product category, translator group, and repeated translations of the same review. Reviews with high ratings and reviews written in more positive language receive feminine forms more often in both languages. All emotion estimates point in the directions gender stereotypes predict, love and sadness toward feminine forms and anger toward masculine ones, but only love in Russian is statistically significant. We find evidence that emotion-gender associations vary across translator groups in Russian, but no clear evidence of such variation in Croatian.

Post-training is widely assumed to reduce the social biases language models inherit from pre-training, yet direct evidence for this assumption remains mixed. In this paper, we examine how gender disparities evolve across the language modeling pipeline. In particular, we study occupational gender bias in OLMo 3, a fully open language model family, across its post-training stages by prompting each checkpoint to generate open-ended narratives about synthetic personas given their occupation and education. We consider two settings, a gender-given setting in which the profile states the persona’s gender explicitly and a gender-assumed setting in which it must be inferred by the model. We find that when gender is stated, post-training increases semantic diversity for male-coded profiles, whereas for female-coded profiles the increase is smaller after SFT and not significant after DPO and RLVR. When gender is assumed by the model, post-training affects narratives about female and male personas similarly. The model’s own gender assumptions, however, show a consistent female-preferential skew relative to United States labor-force demographics, except for heavily male-dominated occupations. Every successive post-training stage monotonically increases the odds of attributing a female label over the base model. A provenance analysis of the supervised fine-tuning corpus finds only a modest 7.7% female-token surplus, too small to explain the size of these downstream effects. The project code is available under: https://github.com/manthangadhia/llm-lifecycle-bias.

Standard English-centric gender bias evaluation provides incomplete coverage of Arabic because Arabic grammar obligates gender choices on verbs, adjectives, pronouns, and nouns. Translating occupation-noun tests or English rewriting benchmarks to Arabic ignores the grammatical forcing that is central to the problem. We argue evaluation for grammatical gender languages requires language-specific extensions starting from the morphology that forces a choice, naming who is affected by the default form chosen, and then assessing whether a system-internal fix helped. Drawing on Arabic Parallel Gender Corpus work, WinoMT, and gender rewriting tools, we offer a protocol for transparent evaluation: name the linguistic condition, including construction type, grammatical person, and number; name the default (the masculine-leaning BLEU gaps); separate linguistic gender from social gender; and do not flatten intersectionality. Arabic evaluation work should be built from Arabic grammar, not adapted from English methods.

Machine translation systems can reproduce gender bias, but most evaluation resources are limited to English as the source language. We introduce DEval-MT, a template-based framework for evaluating gender bias in German-source machine translation, comprising 6,618 sentences across six sentence styles with an intersectional template design that can be extended with additional demographic attributes and support for nonbinary neopronouns. Alongside it, we provide a Python package for translation, alignment, gender prediction, and analysis. Evaluating six different MT systems on seven target languages using this dataset and pipeline, we find evidence of representational harms across multiple demographic groups.

Gender bias in LLMs has been documented across many overt domains, such as colors, emotions, and occupations, while subtler forms rooted in household labor remain understudied. Following Daminger (2019), we distinguish physical household tasks from their cognitive counterparts, which are rarely recognized as labor despite being time-consuming. We adapt the name-generation probing method of Kalhor and Bahrak (2025) to ten household domains, each phrased as both a physical task and its cognitive counterpart, and evaluate six open-source models from three model families in Persian and English. On average, cognitive tasks receive a higher share of feminine names than their physical counterparts in both languages (Wilcoxon signed-rank test); the direction holds in 8 of 10 domains in Persian and 6 of 10 in English. Averaged over models and tasks, skew toward one gender is larger in English than in Persian, in contrast to the overall pattern reported for other domains, but prompts and labeling procedures differ across languages, so we treat this comparison as suggestive. Our results indicate that the cross-linguistic strength of gender bias depends on the domain and model rather than being a fixed property of a language pair.