Large Reasoning Models Are (Not Yet) Multilingual Latent Reasoners
Yihong Liu, Raoyuan Zhao, Hinrich Schütze, Michael A. Hedderich
2026-01-07
Summary
This paper investigates how well large language models (LRMs) can do math problems in different languages, and whether they 'think' through the problem like humans do when they show their work. It turns out these models often figure out the answer *before* fully explaining their steps, suggesting they have a hidden way of reasoning.
What's the problem?
Researchers know LRMs are good at math, and that explaining their steps (called 'chain-of-thought') helps. But recently, it's been discovered that these models sometimes know the answer before finishing the explanation. This means they're doing some reasoning internally, without showing it in the text. The problem this paper tackles is whether this 'hidden reasoning' happens the same way in languages other than English, and if it works equally well for all languages.
What's the solution?
The researchers tested LRMs on math problems in 11 different languages. They didn't let the models finish their explanations completely, instead stopping them at different points. By seeing when the model could still get the right answer, even with an incomplete explanation, they could figure out how much reasoning was happening 'behind the scenes'. They also looked at how the model's internal calculations changed across languages to see if the way it reasoned was consistent.
Why it matters?
This research is important because it shows that while LRMs *can* reason in multiple languages, they don't do it equally well. Languages with more available data for training (like Spanish or French) showed stronger hidden reasoning than languages with less data. Interestingly, the internal reasoning process seemed to be based on how the model learned in English, even when working in other languages. This suggests that improving these models for other languages might involve focusing on how they initially learn to reason in English.
Abstract
Large reasoning models (LRMs) achieve strong performance on mathematical reasoning tasks, often attributed to their capability to generate explicit chain-of-thought (CoT) explanations. However, recent work shows that LRMs often arrive at the correct answer before completing these textual reasoning steps, indicating the presence of latent reasoning -- internal, non-verbal computation encoded in hidden states. While this phenomenon has been explored in English, its multilingual behavior remains largely unknown. In this paper, we conduct a systematic investigation of multilingual latent reasoning in LRMs across 11 languages. Using a truncation-based strategy, we examine how the correct answer emerges as the model is given only partial reasoning traces, allowing us to measure stepwise latent prediction formation. Our results reveal clear evidence of multilingual latent reasoning, though unevenly: strong in resource-rich languages, weaker in low-resource ones, and broadly less observable on harder benchmarks. To understand whether these differences reflect distinct internal mechanisms, we further perform representational analyses. Despite surface-level disparities, we find that the internal evolution of predictions is highly consistent across languages and broadly aligns with English -- a pattern suggesting an English-centered latent reasoning pathway.