Exploring Deep and Wide Search: Insights from the DeepWideSearch Benchmark for LLM-based Agents

This blog post reviews the paper 'A survey of LLM-based Deep Search Agents' (2026), which introduces the DeepWideSearch benchmark—a simulation framework that simultaneously tests LLM-based search agents on deep reasoning and wide information retrieval. The benchmark uses Deep2Wide and Wide2Deep conversion methods to create demanding tasks, revealing that even top-performing LLMs achieved only a 2.39% success rate. The author connects these findings to course topics including A* search algorithms, agent architecture, and multi-step reasoning.