This August 2025 paper introduces Self-Search Reinforcement Learning (SSRL), a novel method that enables Large Language Models (LLMs) to access and utilize their internal knowledge for search-driven tasks, bypassing the need for external search engines like Google or Bing. The research explores how repeated sampling can enhance an LLM's intrinsic search capabilities and investigates the impact of various prompting strategies and training methodologies, including the benefits of information masking and format-based rewards. The paper demonstrates that SSRL-trained models can effectively generalize to real-world search scenarios while often outperforming methods that rely on external search APIs, suggesting LLMs can function as powerful internal knowledge bases for complex queries.Source:https://arxiv.org/pdf/2508.10874

Ver todos los episodios