Language Models Flunk in Multilingual Math Classrooms

Large language models (LLMs) are killing it in math. But only in English. Toss them into a classroom where Sinhala or Tamil are the languages of choice, and watch them crumble. This isn't just a minor glitch, it's a massive gap that could derail tech-driven education in many parts of the world.

Why Sinhala and Tamil?

These languages are staples in South Asian schools, yet they barely get a nod in AI research. Researchers decided to put four top-tier LLMs to the test, using a set of math problems crafted by native speakers. The aim? To see if these models could handle not just English, but also two underrepresented languages: Sinhala and Tamil.

And the outcome? Pretty bleak. While basic arithmetic was a breeze, things got messy with complex problems. You'd think numbers are universal, but LLMs are proving they've some serious blind spots.

Lost in Translation?

Before you blame translation, know this: they didn't translate. Every problem was written separately in each language by people who speak math like a native tongue. So, it's not about bad translation. It's about the models' inability to grasp complex math reasoning beyond English.

The takeaway here's huge. Just because an AI shines in English doesn't mean it's ready to be your global math tutor. Schools looking to integrate LLMs should think twice before rolling them out in non-English settings.

Are We Setting Up Students to Fail?

Deploying these models without proper checks could be setting students up for failure. Imagine trusting an AI tutor that can't even solve your homework in your language. This is a practical nightmare. One that schools can't afford to ignore.

And just like that, the leaderboard shifts. Language-specific evaluation is no longer optional. It's a must. If LLMs are to be the future of education, they need to speak every student's language, literally.

The labs are scrambling to fix this. But the question remains: how long until we see meaningful progress? Because right now, these models might be more of a stumbling block than a stepping stone in multilingual settings.