ZeroSearch: Unlocking LLM Retrieval Capabilities Without Live Search Engines

Introduction: The Dual Challenges in LLM Search Optimization In artificial intelligence development, the retrieval capabilities of Large Language Models (LLMs) fundamentally determine their reasoning quality and generation performance. Current mainstream methods relying on real-time search engines for reinforcement learning training face two critical challenges: 1. Unpredictable Document Quality Existing search engines return documents of varying quality, with high-frequency noise data significantly disrupting training processes. Studies show low-quality documents can reduce model accuracy by 30-40% while creating training instability. 2. Prohibitive API Costs Reinforcement learning requires hundreds of thousands of search requests, with single training sessions potentially exceeding $20,000 using mainstream …