Finally, we recognize the Hawaiian submission to the shared task. Thank you for your contributions!

The second is 7PERFECTION, a dataset for seven Nigerian languages!

The first of our best contribution awards goes to AraPIQA!

We would like to recognize three honorable mention submissions. Great work!

@aliceatkaist.bsky.social joins us to give the final keynote of the day about code switching in multilingual language models!

Join us in hall C3 at posters 137-168 for our in-person poster session. Join us online on Zoom via Underline for our virtual poster session!

Now Pontus Stenetorp shares an oral history of UK-LLM!

@kellymarchisio.bsky.social from Cohere presents “Building Multilingual LLMs in Industry”, sharing insights on training multilinguality at scale!

Preprint: arxiv.org/abs/2510.24081
Dataset: huggingface.co/datasets/mrl...

It’s not too late to get involved! Until early 2026, we will be accepting submissions for languages not already represented in Global PIQA. If you’re interested, please fill out this form and we will contact you with details!
docs.google.com/forms/d/e/1F...

There are seven languages where even the best proprietary LLM scores less than 80% (chance: 50%). Sub-Saharan African languages lag behind Western European languages by ~15%. Thus Global PIQA highlights languages which are very poorly served by large, proprietary models.

The top proprietary models achieve ~90% accuracy, which falls short of human accuracy (~95%). The best open models perform significantly worse, with the best open model performance from Gemma 3 (27B) at 82.4%.

This dataset is created and owned by the contributors, all of whom were offered authorship. We believe this is more fair to annotators and is likely to result in a higher-quality dataset, as it is constructed by the NLP researchers who will use it.

Global PIQA includes subsets for 116 unique language varieties. These cover five continents, 14 language families, and 23 writing systems. Over 50% of examples reference local foods, customs, traditions, or other culturally-specific elements.

sigtyp.github.io/ws2025-mrl.h...

Correct, you can submit ARR papers that already have their reviews by Sep 23. More instructions soon!

TBD on whether the workshop will be hybrid.