Large language models (LLMs) are revolutionizing how we interact with technology, but their impressive capabilities come at a cost: speed. Generating text, especially long passages, can be slow due to the sequential nature of current methods. Imagine waiting for a chatbot to respond or a code completion tool to suggest the next line. It can disrupt the flow and limit real-time applications. But what if we could make LLMs significantly faster, without sacrificing quality? Researchers at Adobe have introduced PLD+, a clever set of algorithms designed to supercharge LLM inference, particularly for tasks where the output is closely related to the input, like code or text editing, summarization, and conversational AI. PLD+ exploits this input-output overlap by intelligently “looking up” parts of the input that are likely to appear in the output. This “draft and verify” approach allows the model to generate larger chunks of text at once instead of word by word. It’s like predictive text on steroids. The key innovation is how PLD+ chooses the best “draft” text spans. It doesn’t just rely on simple string matching. Instead, it uses the LLM’s own internal workings—its “attention” and “hidden states”—to rank potential draft spans by semantic relevance. This is like having an inside track on the model's thought process. The results are impressive. In experiments, PLD+ consistently outperforms existing tuning-free methods and even surpasses some state-of-the-art, fine-tuned approaches in specific tasks. It achieves significant speedups without requiring any model retraining or extra hardware. This means faster code generation, snappier chatbots, and more responsive AI tools overall. While PLD+ shines in input-guided tasks, its performance can dip in scenarios where the output is less dependent on the input, such as open-ended creative writing. Future research could explore ways to adapt PLD+ to a wider range of tasks. The potential is clear: PLD+ offers a practical, out-of-the-box solution for accelerating LLMs and unlocking new possibilities for real-time, interactive AI applications.