Large language models (LLMs) are impressive, but they can still struggle with factual accuracy, often hallucinating information. One solution is to connect them to external knowledge sources like search engines. However, even knowing *what* to search for is a challenge for LLMs. New research explores how to teach these AI models to become better searchers using reinforcement learning. Researchers from Stanford University have developed a technique called "Learning to Retrieve by Trying" (LeReT). The idea is simple: let the LLM experiment with different search queries and learn which ones lead to the most relevant results. Just like humans refine their searches based on what they find, LeReT trains LLMs to do the same. The process involves generating diverse search queries, retrieving documents, and then evaluating their relevance. This feedback loop helps the LLM refine its searching strategy over time. In tests, LeReT significantly boosted the accuracy of retrieved information, leading to more grounded and factual LLM outputs. This is particularly beneficial for complex questions that require multi-hop reasoning, where the LLM needs to synthesize information from multiple sources. While connecting LLMs to the internet empowers them with vast knowledge, teaching them to search effectively is crucial. LeReT represents an exciting step towards making LLMs more reliable and factual by turning them into expert information seekers. Future work might explore how to refine the reward signals used in the learning process and adapt the technique to different types of search engines, potentially revolutionizing how we interact with information online.