HESEIA: Co-Designed Benchmark for Social Biases in LLMs

HESEIA is a community-sourced dataset for auditing and evaluating social biases in Large Language Models. Unlike conventional top-down benchmarks, HESEIA was created through participatory co-design workshops directly within secondary school environments in Argentina and Latin America.

• Scale & Participation: Co-designed with 370 secondary school educators and over 5,000 students across 189 schools, yielding 46,499 sentences capturing intersectional and regionally grounded biases across curriculum subjects.
• Publications & Presentations:
  • Published and presented as an Oral presentation at EMNLP 2025 in Suzhou, China (“HESEIA: A community-based dataset for evaluating social biases in large language models, co-designed in real school settings in Latin America”).
  • Selected for the AI Research & Impact Symposium at the India AI Impact Summit 2026 in New Delhi, India.
• Impact: Bridges the gap between affected educational communities and NLP evaluation methodologies, uncovering biases that standard auto-debiasing prompts fail to mitigate.