ZeroSearch: Incentivize the Search Capability of LLMs without Searching

ZeroSearch is a novel reinforcement learning framework that enables Large Language Models (LLMs) to develop search capabilities without requiring interaction with real search engines. The approach uses an LLM to simulate search engine responses, eliminating expensive API costs while providing control over document quality. Through lightweight supervised fine-tuning and a curriculum-based rollout strategy, ZeroSearch progressively challenges the policy model with increasingly difficult retrieval scenarios, ultimately matching or exceeding the performance of real search engine-based systems.

The method described here would mean that LLM providers would be more independent of the search engine providers and that grounding based on the search engine indices would no longer be necessary, which would mean that SEO would no longer have any influence on the output of the generative AI.