Large language models (LLMs) are impressive, but they can sometimes give wrong answers with unwavering confidence. This "hallucination" problem makes it hard to trust them fully. New research explores a simple yet powerful technique to gauge an LLM's confidence: rephrasing the same question multiple times. The core idea is that if the LLM consistently gives the same answer to slightly different versions of a question, it's likely more confident in its response. Researchers experimented with several rephrasing strategies, from simple synonym substitution to more elaborate expansions of the original question. They found that asking rephrased questions and checking for consistent answers significantly improved the calibration of uncertainty estimates, especially when the model only gave its top prediction (top-1 decoding). Interestingly, this rephrasing method performed almost as well as having access to the model's internal confidence scores (logits), which are usually hidden in closed-source LLMs. This suggests that rephrasing can act as a proxy for true confidence, making it a valuable tool for anyone using LLMs. While top-k decoding, where the model outputs multiple possible answers, generally improves calibration, the study found that combining rephrasing with top-1 decoding offered the best balance between accuracy and confidence estimation. This research opens up exciting possibilities for making LLMs more reliable and trustworthy. Imagine a future where you can quickly rephrase your questions to get a sense of how sure the AI is about its answers – a simple trick with the potential to unlock a deeper level of trust in AI.