Researchers fine-tuned three large-scale AI models to reason in a low-resource language, specifically Greek. They found that accuracy benchmarks showed little improvement, but the models' internal behavior changed significantly. After fine-tuning, the models began to reason in the target language on most items, with improved grammaticality and fluency. However, they also introduced new errors, such as skipping requested formats or leaking into the reasoning channel. Reinforcement learning with verifiable rewards was able to fix some of these defects, but not all. The study proposes six behavioral dimensions for measuring changes in AI models' behavior. --- Why it matters: This research matters because it highlights the limitations of traditional accuracy metrics in evaluating AI models' performance. It also shows that fine-tuning can lead to unintended consequences, such as introducing new errors. Engineers and researchers need to consider these factors when developing and deploying large-scale AI systems. Source: https://arxiv.org/abs/2608.17744