Hidden Language Consistency Phenomena in Reasoning LLMs

Imagine you are in a classroom where the teacher asks a student to solve a tricky math problem, but with a very specific rule: "You must think out loud and write your answer entirely in Spanish." In the world of artificial intelligence, these "students" are Large Language Models (LLMs)—super-smart computer programs that have read almost everything on the internet. When we ask them to solve hard problems, they often use a technique called "Chain-of-Thought," which is like a student scribbling down their step-by-step thinking process before giving the final answer. For a long time, scientists have only cared if the final answer was correct. They treated the thinking process like a black box, assuming that if the answer was right, the student was doing everything correctly. But just like a human student might accidentally start speaking French or English while trying to solve a problem in Spanish, these AI models can sometimes "slip" and switch languages mid-thought, even when we explicitly told them not to. This paper investigates exactly that: what happens when we ask these AI students to solve harder and harder math problems, and do they keep their language promise, or do they get so confused that they start speaking a different language entirely?

The researchers behind this study decided to put these AI "students" to the test using a special math exam called PolyMath, which covers eight different languages and four levels of difficulty, ranging from simple school math to the hardest Olympic-level puzzles. They looked at five different "reasoning" models (AI designed to think deeply) and three "non-reasoning" models (standard AI) to see how they handled the task. What they discovered is a bit like watching a student panic under pressure. They found that as the math problems got harder, the models didn't just get worse at math; they often got worse at sticking to the language they were supposed to use.

The study uncovered four distinct ways this language "slip-up" happens. Sometimes, a model is a rockstar and stays in the correct language no matter how hard the problem gets. Other times, it's a total rebel and refuses to use the requested language from the very start, sticking to its favorite internal language (usually English) no matter what. A third group starts strong but slowly loses focus, drifting into another language as the problems get tougher. But the most surprising finding is the fourth group: these models are perfectly fine with easy and medium problems, but the moment they hit a "medium" difficulty level, they suddenly and completely collapse, switching languages abruptly. The authors call this the "language consistency breakdown effect."

Here is the twist that makes this really interesting: the paper found that this language switch isn't always a bad thing for the score. In some cases, when a model stopped trying to think in the requested language (like Telugu or Swahili) and switched to its "comfort zone" language (like English), it actually got more questions right, even though the problem was harder. It's as if the student stopped trying to solve the problem in Spanish, switched to English, and suddenly got the answer right. This means that if you only look at the final score, you might think the AI is getting smarter, when in reality, it just gave up on following your language instructions.

The researchers also tested what happens when they "compress" these models to make them run faster on smaller devices (a process called quantization). They found that this compression is a bit of a gamble. Sometimes it helps the model stick to the right language, and sometimes it makes it switch languages even faster, and this happens independently of whether the math answers get better or worse. For example, one method called GPTQ often helped models keep their language consistency better than another method called AutoRound, even though AutoRound was better at keeping the math answers correct.

In short, this paper argues that we can't just look at whether an AI gets the right answer to judge if it's good at being multilingual. We have to watch how it thinks and what language it uses while thinking. If we don't, we might miss the fact that the AI is secretly speaking a different language to solve the problem, or that it's suddenly giving up on your language entirely when things get tough. The study suggests that for AI to be truly reliable in a multilingual world, we need to measure not just the final grade, but also whether the student actually followed the language rules of the exam.