Women in AI Research - WiAIR @ NeurIPS 2025
banner
wiair.bsky.social
Women in AI Research - WiAIR @ NeurIPS 2025
@wiair.bsky.social
WiAIR is dedicated to celebrating the remarkable contributions of female AI researchers from around the globe. Our goal is to empower early career researchers, especially women, to pursue their passion for AI and make an impact in this exciting field.
🌱 Why it matters:
Benchmarks like IrokoBench help ensure AI development does not overlook or disadvantage underrepresented languages and communities. (7/8🧵)
December 1, 2025 at 5:07 PM
🌐 Translation boosts accuracy—but shows how heavily models rely on English rather than truly multilingual understanding. (6/8🧵)
December 1, 2025 at 5:07 PM
🔒 Closed-source models like GPT-4o outperform open models across all IrokoBench tasks. (5/8🧵)
December 1, 2025 at 5:07 PM
📉 Large performance gaps appear:
LLMs struggle significantly with African languages, especially those with limited digital resources. (4/8🧵)
December 1, 2025 at 5:06 PM
🧠 AfriXNLI → reasoning
📘 AfriMMLU → knowledge
➗ AfriMGSM → math reasoning
Together, these reveal where AI models fall short beyond English. (3/8🧵)
December 1, 2025 at 5:06 PM
IrokoBench offers a human-translated benchmark for 17 African languages—a major step toward fair and realistic multilingual testing. (2/8🧵)
December 1, 2025 at 5:06 PM
📈 Impact & Gains: Languages with feature data jumped 60.3% (2,724 to 4,366). Overall distance calculation capability grew 9.01%. NLP gains: LANGRANK MT (+6.82%), SemRel2024 (+50%), and a 6.21% RMSE reduction in PROXYLM Unseen. (5/6🧵)
November 26, 2025 at 5:26 PM
The SoftImpute algorithm is the top performer. It achieved an F1 score of 0.7980 on union data and the lowest error (RMSE 0.2883) on average-aggregated data. All imputation methods outperform the mean baseline. (4/6🧵)
November 26, 2025 at 5:26 PM
It adds 133 morphological features and customizable distance metrics (union/average aggregation, angular/cosine distance). It also provides confidence scores based on data quality. Three algorithms are available for missing data. (3/6🧵)
November 26, 2025 at 5:25 PM
🔍 What's New? URIEL+ integrates five new databases and uses Glottocodes to support low-resource languages (LRLs) like Eskimo Pidgin and Singlish. This greatly expands language inclusion beyond standard ISO codes. (2/6🧵)
November 26, 2025 at 5:25 PM
👇 Watch the trailer + subscribe so you don’t miss the full episode!

youtu.be/Bh3bT-r4aH8
youtu.be
November 17, 2025 at 5:03 PM
Annie's research focuses on language diversity, multilinguality, low-resource languages and multicultural fairness in AI systems.
November 14, 2025 at 4:01 PM