Large language models (LLMs) have taken the world by storm, demonstrating impressive abilities in writing, translation, and even coding. However, these powerful AI systems sometimes stumble, generating convincing yet incorrect information—a phenomenon known as 'hallucination.' How can we make LLMs more trustworthy? New research explores a promising solution: uncertainty-aware fine-tuning. Just like humans express varying levels of confidence in their answers, researchers are teaching LLMs to recognize when they might be wrong. This involves training the models to provide not just an answer, but also an estimate of their own uncertainty about that answer. Imagine an LLM suggesting a medical diagnosis alongside a confidence score. This extra information allows users to judge the reliability of the LLM's output and proceed with caution if the uncertainty is high. The research introduces a novel approach called 'uncertainty-aware causal language modeling,' which essentially encourages the LLM to express high uncertainty when it's likely to make a mistake and low uncertainty when it's confident in its response. Experiments on various question-answering datasets show that this method significantly improves the LLM's ability to flag potentially incorrect answers, detect out-of-domain questions (topics it hasn't been trained on), and even strategically decide when to abstain from answering altogether. This represents a significant step toward building more reliable and trustworthy LLMs for critical applications where accurate information is paramount. While the current method focuses on accessing the inner workings of the model (white-box setting), future research aims to extend this approach to scenarios where such access is limited. The ultimate goal is to equip LLMs with a robust sense of their own limitations, making them more valuable and reliable tools in various fields.