Large language models (LLMs) are known for their impressive text generation abilities, but they've also shown promise in another area: creating sentence embeddings. These embeddings are essentially numerical representations of sentences that capture their meaning, useful for tasks like search, clustering, and comparing text similarity. However, getting high-quality sentence embeddings from LLMs hasn't been straightforward. A new research paper introduces a clever trick called 'Token Prepending' (TP) that significantly boosts the quality of LLM-generated sentence embeddings *without any extra training*. The key insight? LLMs, due to their design, sometimes struggle to fully grasp the meaning of a sentence because they process words sequentially, missing crucial backward references. TP solves this by strategically inserting a special token, , into the input. This token acts as a placeholder, allowing the model to build a richer understanding of the sentence’s meaning as it processes each word. In essence, the token helps the model ‘look back’ at the whole sentence, even though it reads words from left to right. The results? Experiments across various LLMs, including LLaMA2 and Qwen2, show that TP consistently improves the quality of sentence embeddings, leading to better performance in tasks like semantic text similarity and transfer learning. What makes this technique particularly appealing is its efficiency and ease of use. TP doesn’t require any changes to the model's architecture or additional training data. It’s a simple adjustment that can be applied to any LLM, unlocking its hidden potential for sentence understanding. This research opens exciting doors for using LLMs in a broader range of applications. By simply prepending a token, we can improve sentence embeddings and boost performance in numerous downstream tasks, pushing the boundaries of what LLMs can achieve.