Skip to content

ZeroSearch

Train LLM search abilities without a live search engine

Visit official site

What ZeroSearch is

The framework is described as a reinforcement learning approach that builds search abilities in large language models without interacting with a real search engine. Supervised fine-tuning turns an LLM into a retrieval module generating relevant and irrelevant documents, with a curriculum rollout mechanism for reasoning.

What you can do with it

  • Optimise retrieval ability through reinforcement learning
  • Apply supervised fine-tuning to improve the model
  • Skip interaction with a real search engine
  • Use curriculum rollout to build reasoning step by step

Getting started

  • Create a Conda environment and install the dependency packages
  • Download the training dataset and the simulated LLM
  • Start the local simulation server
  • Set the Google Search API key

Model notes

  • The framework fits LLMs of various sizes
  • Different reinforcement learning algorithms are supported
  • The material states zero API cost and performance ahead of models based on real search engines

Pros & cons

✓ What we like

  • No interaction with a real search engine is needed
  • Fits LLMs of various sizes and different reinforcement learning algorithms

! What to watch out for

  • Setup requires a Conda environment and dependency installation
  • The documented steps include downloading a training dataset and a simulated LLM

FAQ

Does it need a real search engine?

The material states the framework builds search abilities without interacting with an actual search engine.

What does the setup involve?

The listed steps cover a Conda environment, training data and a simulated LLM, a local simulation server, and a Google Search API key.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review