Large language models (LLMs) are impressive, but they sometimes struggle with honesty. They can confidently assert incorrect information or avoid answering questions they actually *do* know, just not with complete certainty. This 'knowledge boundary' problem limits their reliability. New research introduces UAlign, a framework that teaches LLMs to be more truthful by leveraging their own uncertainty. UAligne explicitly incorporates two uncertainty measures – confidence scores and semantic entropy – into the LLM's training process. Think of it like giving the model a built-in 'doubt meter.' Confidence scores represent how sure the LLM is about an answer, while semantic entropy captures the dispersion of different possible responses. By feeding these measures back to the model, UAlign helps it distinguish between what it knows well, what it's less certain about, and what it truly doesn't know. This allows the model to confidently answer questions within its knowledge boundary, even if some uncertainty exists, while also admitting when it's stumped. Experiments show that UAlign significantly improves LLM honesty and reliability across diverse knowledge domains. It's particularly effective in generalizing to new, unseen questions. This suggests that explicitly modeling uncertainty could be key to making LLMs more trustworthy and reliable sources of information. While computationally intensive, the initial results of UAlign offer a compelling direction for LLM training, highlighting the importance of acknowledging and managing uncertainty in the pursuit of truly intelligent AI.