Papers by Mahammed Kamruzzaman
BanStereoSet: A Dataset to Measure Stereotypical Social Biases in LLMs for Bangla (2025.findings-acl)
Copied to clipboard
| Challenge: | ***BanStereoSet*** is a dataset designed to evaluate stereotypical social biases in multilingual LLMs for the Bangla language. |
| Approach: | They propose to localize the content from StereoSet, IndiBias, and kamruzzaman-etal's datasets to capture biases prevalent within the Bangla language. |
| Outcome: | The proposed dataset consists of 1,194 sentences spanning 9 categories of bias: race, profession, gender, ageism, beauty, beauty in profession, region, caste, and religion. |
Seeing Race, Feeling Bias: Emotion Stereotyping in Multimodal Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Emotion stereotypes are also tightly tied to race and skin tone, but previous studies have overlooked this dimension. |
| Approach: | They propose a multimodal study of racial, gender, and skin-tone bias in emotion attribution . they evaluate four open-source MLLMs using 2.1K emotion-related events . |
| Outcome: | The proposed study examines four open-source MLLMs using 2.1K emotion-related events paired with 400 neutral face images across three different prompt strategies. |
Exploring Changes in Nation Perception with Nationality-Assigned Personas in LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Using personas, LLMs are custom-made to meet specific user needs. |
| Approach: | They assign 193 different nationality personas to five LLMs and examine how evaluations of different nations change when LLM users are assigned specific nationality persons. |
| Outcome: | The nationality personas of five LLMs are assigned to different nations and their evaluations change. |
Investigating Subtler Biases in LLMs: Ageism, Beauty, Institutional, and Nationality Bias in Generative Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in language generation models can be used to assist users in a variety of tasks, but there are risks associated with introducing LLM biases into consequential decisions. |
| Approach: | They propose to use a template-generated dataset to measure subtler correlated decisions that LLMs make between social groups and unrelated positive and negative attributes. |
| Outcome: | The proposed model can be used to evaluate progress in more generalized biases and extend the benchmark with minimal human annotation. |
The Impact of Name Age Perception on Job Recommendations in LLMs (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing studies have examined implicit age-related name bias in LLMs . older-sounding names are favored for senior roles, while younger-sounders are linked to youth-dominant jobs . |
| Approach: | They analyze six LLMs and 117 American names categorized by perceived age across 30 occupations . older-sounding names are favored for senior roles, while younger-sounders are linked to youth-dominant jobs . |
| Outcome: | The proposed model based on six LLMs and 117 American names shows that older-sounding names are favored for senior roles, while younger-sounders are linked to youth-dominant jobs. |
“Global is Good, Local is Bad?”: Understanding Brand Bias in LLMs (2024.emnlp-main)
Copied to clipboard
| Challenge: | a recent study examined social biases in LLMs but brand bias has received little attention. |
| Approach: | They examine the behavior of LLMs in the market place by analyzing a brand-based dataset . they find a consistent pattern of brand bias in this space . |
| Outcome: | The proposed model favors established global brands while marginalizing local ones . the proposed model could boost local brand preference in LLM outputs in specific contexts . |