CS Paper Radar

How Humans and LLMs Read Gender into "Gender-Neutral" Physical Descriptions

Yingjia Wan, Lin Lin, Elisa Kreiss

cs.CL cs.AIcs.CYcs.HC ·submitted 2026-09-14 21:29 UTC · v2
publishedcode/datahot direction

Abstract

When foundation models describe people, recent work in AI fairness, accessibility, and ethics recommends avoiding inferred identity labels (e.g., "she", "his") in favor of seemingly "objective" physical descriptions (e.g., "short hair", "a defined jawline"). Yet whether such descriptive language achieves gender-neutral communication remains an open empirical question. To study this, we introduce GAPA (Gender Associations of Physical Attributes), a dataset of 316 common physical attributes drawn from diverse sources, paired with 14,706 gender-association ratings from 304 US-based annotators. Results show that physical descriptions carry structured and graded gender associations among readers, with more consistent and distinctive associations for women and men than for non-binary identities. Next, we evaluate 16 LLMs across model families, sizes, and post-training variants against human ratings. The models partially recover human associations but exhibit systematic alignment biases, including compressed rating distributions, weaker alignment for associations with men, and asymmetric abstention that disproportionately targets the non-binary category. Finally, we release the best-performing proxy model trained to predict humans' gender associations of descriptive language and demonstrate its utility through a sociolinguistic analysis of character descriptions in LitBank. Together, our findings provide the first empirical evidence that seemingly "objective" physical descriptions can retain systematic gender associations in human interpretation, and uncover systematic patterns of model-human misalignment. This challenges the assumption that replacing explicit gender labels with physical descriptions necessarily yields gender-neutral communication, and highlights downstream challenges in using such descriptions to communicate subjective identity categories in human-AI interaction.

Comment
The dataset and code are available at https://github.com/Yingjia-Wan/GAPA, and the predictor model is released at https://huggingface.co/alisa-yingjia-wan/gapa-predictor-olmo2-7b
Journal ref
In Proceedings of Third Conference on Language Modeling (COLM), 2026

Closest papers

Nearest neighbours by meaning across the whole index.

GRAP: Gender Rubric-Aware Probing for Analyzing Gender Bias in LLMs

cs.AI ·2026-11-09 ·Seona Seo, Geonhee Ahn, Hyunsoo Cho, Sookyung Kim
similarity 0.66
Abstract · 3 pages. Accepted at the ECML-PKDD 2026 Workshop on Bias and Fairness in AI

As LLMs increasingly mediate everyday interactions, subtle gender biases in open-ended responses pose growing societal risks. Existing probing methods have two key limitations: they rely on fixed-format tasks, such as pronoun resolution or occupation association, that alignment-tuned models can easily handle; and they compare raw group-level rubric scores, ignoring per-dimension variance and obscuring which bias dimensions reliably distinguish gender groups. We introduce GRAP (Gender Rubric-Aware Probing), a framework that uses a lightweight logistic probe to identify the most discriminative bias dimensions from responses scored across 10 rubric dimensions, operating task-agnostically on open-ended instruction following.

0.17

No-Shot Text-to-Speech: Limitations of Zero-Shot TTS and its Evaluation Methods in Representing Queer and Transgender Voices

cs.SD ·2026-11-16 ·Juliana Francis, Robin Netzorg, Joakim Gustafson, Éva Székely
top venue similarity 0.65
Abstract · Proceedings of Interspeech 2026

Gender expansive voices remain underrepresented in speech datasets, which could cause challenges with generalizability in zero-shot Text-To-Speech (TTS) systems. We evaluate six zero-shot TTS models on two datasets: one containing non gender expansive voices (N-GE) and one with gender expansive (GE) voices from self-identified speakers. Human ratings found that 2 of the 6 models synthesize less similar GE voices compared to N-GE voices, and 3 have improved performance for GE voices. Automated evaluations yield conflicting results, with speaker similarity and automated mean opinion score predictors rating performance of GE voices inconsistently between models and human ratings. This highlights a potential blind spot in evaluation pipelines, emphasizing a need for more representative data and metrics that better capture human perceptual judgments. Our findings stress the need to include transgender, queer, and minority communities in creation of better datasets and model training.

0.39

Beyond Attribute-Centric Fairness: Modelling and Mitigating Unified Bias Representation in Omni-modal Large Language Models

cs.LG ·2026-11-17 ·Runqi Lin, Suqin Yuan, Zihao Fu, Chris Russell
hot direction similarity 0.65
Abstract

Omni-modal large language models (OLLMs) provide joint multimodal understanding and generation capabilities, making them well-suited to complex environments and promising backbones for next-generation AI. Despite these advances, the harmful stereotypes encoded in such models manifest in increasingly diverse forms, leading to broader and more severe societal consequences. Existing debiasing methods typically impose separate fairness objectives for each sensitive attribute and modality to minimise demographic-dependent variations. However, in omni-modal settings, this attribute-centric paradigm struggles to jointly address diverse social biases with distinct fairness objectives and to generalise beyond predefined attributes. To this end, we propose OmniDebias, one of the first methods for promoting fairness in OLLMs, which employs a bias-centric formulation to extract and mitigate a shared bias representation across attributes and modalities while preserving attribute-related knowledge. Specifically, we exploit activation inconsistencies of a given attribute (e.g., gender) across semantic contexts (e.g., occupations) to disentangle bias from legitimate knowledge. We further consolidate multiple extracted biases into a shared representation, formulating a unified objective that captures the underlying structure across social biases and generalises to unseen attributes. Extensive experiments demonstrate that eliminating unified bias with both post-hoc editing and fine-tuning can consistently improve fairness across diverse attributes and modalities in OLLMs while maintaining favourable utility.

0.28

LLMs Cannot Be Neutral: Gender Sensitivity Across Candidates, Evaluators, and Models

cs.CL cs.HC ·2026-11-02 ·Danial Amin, Joni Salminen, Bernard J. Jansen
hot direction similarity 0.65
Abstract · This manuscript is under review at FAccT'2027

Large language models (LLMs) are increasingly used to evaluate candidates for professional roles, raising concerns about their sensitivity to demographic identities. We examine how candidate and evaluator gender assignments affect decisions by LLM-as-a-Judge (LaJ) models and decision language models (DLMs). In a hypothetical city council voting task, eight models evaluated 240 synthetic professional narratives from NVIDIA's US and French Nemotron-Personas datasets under independently varied gender assignments, producing 30,720 judgments. Explicit candidate gender labels increased Yes-vote probability by 1.05-1.73 percentage points relative to unspecified gender, while only the nonbinary evaluator assignment showed a significant increase (1.26 points). Identity matching showed no significant overall association. Yes-vote rates varied across models (81.3-97.2%), and 2.5-3.3% of decisions reversed relative to the unspecified-identity baseline. Under direct substitutions between explicit gender identities, DLMs exhibited fewer decision reversals than LaJ models (1.18% vs. 3.38%), indicating lower observed decision-level gender sensitivity among the tested DLMs. These findings demonstrate that gender sensitivity depends on candidate and evaluator identities and the evaluation model, while highlighting the importance of examining individual decision consistency alongside aggregate differences.

0.31

Algorithmic Gender Prediction Is Illegitimate, But Gender Imputation Can Yield Valid Measurements

cs.CY ·2026-08-13 ·Evan Dong, Angelina Wang
publishedposition similarity 0.64
Abstract · Accepted to the 9th AAAI/ACM Conference on AI, Ethics and Society (AIES-26)

Machine learning ethics researchers and critical HCI scholars have argued that algorithmically predicting gender is wrong. At the same time, other researchers rely on predicted gender labels to study gender disparities and develop algorithmic fairness techniques. How do we reconcile these two seemingly contradictory intuitions? We differentiate two ways gender prediction may be wrong: being illegitimate, thereby contributing to harm; and being invalid, thereby producing unusable measurements. Our analysis translates arguments against gender prediction into these terms of legitimacy and validity and shows how gender imputation applied for fairness purposes can be illegitimate yet still yield valid disparity measurements. We clarify this bind by drawing upon transfeminist literature to distinguish sexism that targets women and femininity from sexism that targets transgender and nonbinary people. While gender imputation can produce valid measurements for the former, it is illegitimate and harmful for the latter. We argue that practitioners should deploy gender imputation only when it would achieve anti-discrimination benefits that cannot be achieved through other reasonable means, while harms are minimized to the extent possible. We examine this tension in three case studies: auditing gender bias in generative image models, measuring gender disparities in film, and imputing gender from personal names. By disentangling legitimacy from validity, and differentiating these two forms of sexism, we show how debates over gender prediction have conflated distinct concerns, obscuring both the settings in which gender imputation can support fairness efforts and the harms towards transgender and nonbinary people that it fundamentally cannot capture. We conclude by recommending the development of more inclusive methods that address all kinds of sexism.

0.17

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

cs.CL ·2026-11-02 ·Zihao Sun
similarity 0.63
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

Analyzing Occupational Gender Bias across the LLM Training Lifecycle

cs.CL ·2026-10-16 ·Manthan Gadhia, Alexander Hoyle, Karolina Stańczak
code/data similarity 0.63
Abstract · Submitted to 7th Workshop on Gender Bias in Natural Language Processing @AACL-IJCNLP 2026

Post-training is widely assumed to reduce the social biases language models inherit from pretraining, 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.

0.37

Portrait Fitness as a Pretext: Probing Implicit Demographic Bias in VLMs

cs.CV cs.CY ·2026-11-16 ·Alexandre Fournier-Montgieux, Akram Elbouanani, Aboubacar Tuo, Adrian Popescu, Hervé Le Borgne +1 more
published similarity 0.62
Abstract · Accepted at WACV 2027

We audit face-related bias in VLMs using BALPORTRAITS, a new balanced dataset of 9,000 fictional portraits across ethnicity, gender, and age segments. Sixteen VLMs estimate portrait fitness scores for occupations, brands, activities, and banking or insurance products within an advertising-casting framework designed to avoid direct inference of protected attributes. Item-level estimations are aggregated with scenario-specific prestige scores to compare rankings across demographic groups. While the magnitude and direction of these biases vary across scenarios, model families, and demographic dimensions, with age, ethnicity, and gender contributing differently depending on the task, we find demographic dispersion across all evaluated models, with a majority of pairwise segment differences reaching statistical significance. This study introduces a controlled protocol to assess implicit face-related associations and demonstrates that single-scenario or single-attribute audits are insufficient.

0.50