Large language models (LLMs) are impressive, but adapting them to specific tasks often requires resource-intensive fine-tuning or clever prompt engineering. What if there was a faster, simpler way to boost their performance? Researchers have unveiled a promising new technique called Reference Trustable Decoding (RTD), a training-free method that lets LLMs tap into external knowledge without complex fine-tuning or lengthy prompts. RTD works by creating a 'reference datastore' from task-specific examples. When the LLM encounters a new problem, RTD quickly finds the most relevant references in the datastore and uses them to refine the LLM's output, essentially giving it a boost of task-specific knowledge. This approach is not only faster than traditional fine-tuning, it also avoids the memory bottlenecks of in-context learning, where providing lots of examples can slow down processing. In tests across various LLMs and benchmarks, RTD achieved comparable or even better performance than existing methods, particularly in language understanding tasks like question answering. The results are particularly striking in scenarios requiring complex reasoning, such as knowledge injection from Wikipedia. RTD demonstrated significant accuracy gains while using considerably shorter prompts, which translates to faster processing. While promising, RTD faces challenges, especially with larger datastores. Managing these efficiently is key to unlocking its full potential. Future research will focus on automatically identifying and eliminating redundancies in these datastores to enhance performance without sacrificing efficiency. RTD represents a significant leap forward in making LLMs more adaptable and efficient, paving the way for broader applications across various fields.