The United Nations’ Sustainable Development Goals (UN SDGs) prioritise inclusive and fair employment. However, AI-powered recruitment tools—particularly Large Language Models (LLMs)—raise concerns about potential demographic bias. This paper presents a controlled synthetic dataset and methodology to measure how sensitive attributes (e.g., race, gender, age) influence candidate rankings and pairwise comparisons in LLM-based hiring pipelines. Specifically, we generated a balanced dataset of 1,000 synthetic candidate profiles (each including a cover letter) and evaluated it using 28 frontier LLMs, including proprietary (e.g., OpenAI GPT, Gemini, Grok, Claude) and opensource (e.g., Llama, GigaChat) models. Synthetic data eliminates real-world demographic/occupational confounders, ensuring observed disparities reflect only LLMs’ intrinsic behaviour. Results show professional attributes (e.g., skills, experience) are primary ranking drivers, with 76%–80% statistically significant; however, 8%–9% of demographic attributes exhibit persistent, significant biases across multiple LLMs.We develop a “bias map” quantifying LLM performance, emphasising that mitigating even minor biases in automated hiring is critical to avoid perpetuating employment inequities and uphold the UN SDGs’ inclusive vision.

Medical visual question answering (Med-VQA) is to answer medical questions based on clinical images provided. This field is still in its infancy due to the complexity of the trio formed of questions, multimodal features and expert knowledge. In this paper, we tackle, a ’myth’ in the Natural Language Processing area - that unimodal bias is always considered undesirable in learning models. Additionally, we study the effect of integrating a novel dynamic attention mechanism into such models, inspired by a recent graph deep learning study.Unlike traditional attention, dynamic attention scores are conditioned on different query words in a question and thus enhance the representation learning ability of texts. We propose that some questions are answered more accurately with a reinforcement of question embedding after fusing multimodal features. Extensive experiments have been implemented on the VQA-RAD datasets and demonstrate that our proposed model, reinforCe unimOdal dynamiC Attention (COCA), outperforms the state-of-the-art methods overall and performs competitively at open-ended question answering.