Researchers have found that a common technique for training large language models on mathematical reasoning tasks, called reinforcement learning with verifiable rewards (RLVR), can be biased against certain languages. An exact-match verifier used in RLVR can mistakenly reject correct answers from some languages due to formatting or script differences. The study proposes a framework for auditing multilingual RLVR rewards and identifies a 'cross-lingual selection bottleneck' that hinders accurate answer selection across languages. This issue is particularly pronounced when using models like Qwen3-8B, which shows a false-negative rate of 64% on Japanese answers compared to English or Chinese. The study suggests that multilingual RLVR rewards should be audited and optimized separately for each language and answer interface. --- Why it matters: This matters because it highlights the need for more nuanced approaches to training large language models on mathematical reasoning tasks, particularly in multilingual settings. It also underscores the importance of auditing and optimizing model performance across different languages and interfaces. Source: https://arxiv.org/abs/2608.20362