Large language models (LLMs) are impressive, but they sometimes confidently give wrong answers, a problem called "hallucination." One solution is to teach them when to abstain, or say "I don't know." However, current techniques mostly work for English. New research tackles how to teach LLMs to abstain in multiple languages. The challenge is that these models' knowledge and reasoning abilities are often weaker in languages other than English. Researchers found that simply translating English abstention methods doesn't work well – there's a significant performance drop in other languages, particularly under-resourced ones. Their solution? Multilingual feedback. The LLM generates possible answers, then provides feedback on its own answers *in multiple related languages*. This cross-linguistic reflection helps identify gaps and inconsistencies across different perspectives. For example, if an LLM is asked a question in Bengali, it might also generate feedback in Hindi, Marathi, and Nepali. This helps catch errors an English-only or even just Bengali-only approach might miss. It turns out related languages often reveal subtle cultural and contextual differences that highlight weaknesses in the model's understanding. The results? This method significantly outperforms existing techniques, especially for lower-resourced languages. It not only improves accuracy but also leads to more equitable performance across languages. Interestingly, the researchers found that cultural factors strongly influence how LLMs abstain. Questions on Western-centric topics like U.S. history were more likely to cause inconsistencies across languages, highlighting the need for more culturally diverse training data. This research shows promising progress in making LLMs more reliable and trustworthy across languages, opening doors to broader global access to accurate information. Future research could explore how to optimize language selection for feedback and tailor strategies for specific languages and cultural contexts.