SDGBiasBench: Benchmarking and Mitigating Vision--Language Models' Biases in Sustainable Development Goals
Researchers have introduced SDGBiasBench, a new benchmark designed to evaluate and mitigate biases in vision-language models (VLMs) concerning the Sustainable Development Goals (SDGs). The benchmark includes over 500,000 multiple-choice questions and 50,000 regression tasks, revealing that current VLMs often rely on SDG-specific priors rather than visual evidence. To address this, the team developed CADE, a training-free method that improves model accuracy by up to 25% and reduces estimation errors by 12 points. AI
IMPACT Introduces a new evaluation framework and debiasing technique for AI systems focused on sustainable development.