ZeroSearch: Fostering LLM search capabilities without real search engine

🔥Zero-Cost Training for Simulated Search
ZeroSearch offers a lightweight supervised fine-tuning method that transforms large language models (LLMs) into simulated search engines. This approach enables controllable document generation while avoiding the instability and high API costs associated with real search engines.

🔥Curriculum Reinforcement Learning for Enhanced Reasoning
The project introduces a curriculum training strategy that progressively increases the noise level in generated documents. This method forces the model to enhance its reasoning abilities in more challenging retrieval scenarios, creating a learning process that moves from easy to difficult.

🔥Efficient Training, Superior Performance
Experiments demonstrate that a 3-billion-parameter simulator can activate LLM search capabilities. A 7-billion-parameter simulator performs comparably to real search engines, and a 14-billion-parameter simulator even surpasses real search engines, all while reducing training costs by a remarkable 88%.

Видео ZeroSearch: Fostering LLM search capabilities without real search engine канала Tongyi Lab