Key Points

1. Information seeking and integration is a complex cognitive task that consumes significant time and effort. Search engines have reshaped how information is sought, but often fail to align with complex human intentions.

2. Inspired by the progress of Large Language Models (LLMs), recent works have attempted to solve the information-seeking and integration task by combining LLMs and search engines, but face three key challenges: 1) complex requests cannot be accurately retrieved by search engines in a single query, 2) relevant information is spread across multiple web pages with massive noise, and 3) the volume of web content can exceed the context length of LLMs.

3. The paper introduces MindSearch, a multi-agent framework consisting of a WebPlanner and WebSearchers, to mimic the human cognitive process for web information seeking and integration.

4. The WebPlanner models the problem-solving process as a dynamic graph construction, decomposing the user query into atomic sub-questions and progressively extending the graph based on search results.

5. The WebSearcher performs hierarchical information retrieval to extract valuable information for the WebPlanner, improving the efficiency of information aggregation.

6. The multi-agent design of MindSearch enables parallel information seeking and integration from over 300 web pages in under 3 minutes, compared to 3 hours of human effort.

7. MindSearch demonstrates significant improvements in response quality in terms of depth and breadth, outperforming ChatGPT-Web and Perplexity.ai on both closed-set and open-set QA problems.

8. Responses from MindSearch using the open-source InternLM2.5-7B model are preferred by human evaluators over the proprietary ChatGPT-Web and Perplexity.ai applications.

9. The multi-agent framework of MindSearch provides a simple yet effective solution for synergizing the web information retrieval capabilities of search engines and the reasoning power of LLMs.

Summary

The research paper introduces a novel LLM-based multi-agent framework called MindSearch, designed to address the challenges of information seeking and integration using a combination of search engines and Large Language Models (LLMs). The paper identifies three main challenges in the current approaches, including the inability to accurately retrieve complex requests, the spread of information across multiple web pages with noise, and the limitation of LLM context length in dealing with large amounts of web content. MindSearch aims to mimic the human mind in web information seeking and integration by instantiating a multi-agent framework consisting of a WebPlanner and WebSearcher.

Multi-Agent Design of MindSearch
The WebPlanner models the human mind as a dynamic graph construction process, decomposing the user query into atomic sub-questions and extending the graph based on search results from WebSearcher. WebSearcher performs hierarchical information retrieval with search engines and collects valuable information for WebPlanner. The multi-agent design of MindSearch enables parallel information seeking and integration from a large number of web pages in a short time period, showing significant improvements in response quality compared to existing approaches.

The Significance of MindSearch
The paper highlights the necessity of information seeking and integration in all walks of life, the limitations of current search engines, and the potential of Large Language Models (LLMs) to address these challenges. It discusses the significance of MindSearch as a competitive solution for AI-driven search engines, providing a detailed overview of the framework's architecture, components, and operational processes. The WebPlanner functions as a high-level planner, orchestrating the reasoning steps and coordinating other agents, while the WebSearcher acts as an RAG (Retrieve-and-Generate) agent with internet access for web browsing and information summarization.

Effectiveness of MindSearch
The paper also evaluates the effectiveness of MindSearch using closed-set and open-set Question Answering (QA) tasks, demonstrating substantial improvements in response quality, depth, breadth, and factuality. The empirical findings show that MindSearch outperforms existing applications and competitive solutions, indicating its potential to provide a highly competitive solution for AI-driven search engines. The paper also discusses the integration of web search capabilities into LLMs, addressing context management and context transfer across multiple agents. Finally, the paper discusses the broader implications and future research directions for multi-agent frameworks in solving complex cognitive tasks.

Overall, the research paper presents MindSearch as an innovative and effective approach to information seeking and integration, demonstrating its potential to achieve significant improvements in response quality and competitiveness in AI-driven search solutions. The detailed architecture, processes, and empirical validations contribute to a comprehensive understanding of the MindSearch framework and its implications for future research in the field.

Reference: https://arxiv.org/abs/2407.20183