DeepRetrieval introduces a novel reinforcement learning approach that trains large language models to generate optimized search queries without requiring expensive supervised learning or labeled data.

• Trains LLMs through trial and error to generate queries that yield better search results
• Eliminates the need for hand-labeled training data or complex distillation techniques
• Demonstrates effectiveness across multiple search environments including commercial search engines
• Improves search precision while reducing computational costs

Security Implications: This research reveals how LLMs can be used to systematically optimize queries that extract specific information from search engines, potentially bypassing intended information access controls or manipulating search rankings.

DeepRetrieval: Hacking Real Search Engines and Retrievers with Large Language Models via Reinforcement Learning