Search-o1
Search-o1 is an agentic framework designed to enhance Large Reasoning Models (LRMs) by integrating web search capabilities directly into their reasoning processes. Accepted at EMNLP 2025, this system addresses a critical limitation of current LRMs: knowledge insufficiency during long chain-of-thought reasoning. By enabling models to autonomously search for real-time information, retrieve relevant facts, and verify their own reasoning steps, Search-o1 significantly reduces uncertainties and errors in complex problem-solving tasks. The framework is particularly effective for domains requiring up-to-date or specialized external knowledge, such as advanced mathematical reasoning, scientific query resolution, and answers on benchmark datasets like AIME24, AMC23, and GPQA. The project includes code implementation for researchers to apply search-enhanced reasoning to their own models, with future plans to support diverse backbone architectures like DeepSeek-R1 and Sky-T1, expand tool usage to include code interprete