Large language models often struggle with factuality, sometimes hallucinating incorrect information. This issue arises from various factors, including biased or incomplete training data. Factuality, the ability to generate truthful content, is crucial for reliable LLM applications. SLED, a new decoding method, aims to improve factuality without external knowledge bases. SLED leverages information from all layers of the LLM, not just the final layer, to refine its predictions. It calculates token probabilities using earlier layers, assigning weights to each for a more accurate output. Experiments on multiple tasks and benchmarks show SLED improves factual accuracy across different LLMs. For example, it can correct math errors or choose the correct answer to a multiple-choice question. SLED is easily implemented, compatible with various LLMs, and can be combined with other methods. Its primary tradeoff is a minimal increase in inference time compared to alternatives. SLED demonstrates state-of-the-art accuracy improvements without requiring extensive fine-tuning. Future work may involve combining SLED with supervised fine-tuning and applying it to other tasks.