Large language models (LLMs) are impressive, but they sometimes struggle with accuracy, making up facts or hallucinating. This poses a serious problem for applications requiring reliability. New research explores a clever method called Self Logits Evolution Decoding (SLED) to tackle this issue. Instead of relying on external fact-checking or retraining the model, SLED taps into the LLM's existing knowledge. The idea is to compare the model's final output with its internal, earlier processing stages. By examining these differences, SLED tries to identify and correct inconsistencies, essentially guiding the model to refine its own answers. The researchers tested SLED across different models, from the smaller LLaMA 2 and Gemma models to the larger 70B parameter LLaMA 2 and the complex Mixture of Experts (MoE) models like Mixtral. Across these various setups, SLED demonstrated a boost in factuality. In tasks involving multiple-choice questions, open-ended generation, and complex reasoning problems, SLED consistently outperformed traditional methods and prior techniques like DoLA. Impressively, SLED reduces the tendency of models to simply refuse to answer difficult questions, leading to more informative responses. While promising, SLED isn't a silver bullet. It adds some computational overhead, though not substantial. Also, like all current solutions, it doesn't perfectly eliminate errors. Nevertheless, it offers a novel, cost-effective path towards more truthful and reliable AI, hinting at the potential of self-correction within these powerful models. Future work could explore combining SLED with other techniques for enhanced accuracy, opening doors to more trustworthy and useful AI applications.