ManuSearch: Democratizing Deep Search in Large Language Models with a Transparent and Open Multi-Agent Framework (2505.18105v1)

Abstract: Recent advances in web-augmented LLMs have exhibited strong performance in complex reasoning tasks, yet these capabilities are mostly locked in proprietary systems with opaque architectures. In this work, we propose \textbf{ManuSearch}, a transparent and modular multi-agent framework designed to democratize deep search for LLMs. ManuSearch decomposes the search and reasoning process into three collaborative agents: (1) a solution planning agent that iteratively formulates sub-queries, (2) an Internet search agent that retrieves relevant documents via real-time web search, and (3) a structured webpage reading agent that extracts key evidence from raw web content. To rigorously evaluate deep reasoning abilities, we introduce \textbf{ORION}, a challenging benchmark focused on open-web reasoning over long-tail entities, covering both English and Chinese. Experimental results show that ManuSearch substantially outperforms prior open-source baselines and even surpasses leading closed-source systems. Our work paves the way for reproducible, extensible research in open deep search systems. We release the data and code in https://github.com/RUCAIBox/ManuSearch

Deep Search in LLMs: An Overview of ManuSearch

The paper introduces ManuSearch, a transparent and modular multi-agent framework designed for deep search tasks in LLMs. With the increasing complexity of queries that LLMs encounter, the necessity for systems that can reason over multi-step processes and integrate real-time web information has become apparent. ManuSearch addresses the limitations seen in proprietary systems such as Perplexity's Sonar Reasoning Pro and OpenAI's GPT-4o Search Preview, which are not open-source and lack transparency in their architectures.

Framework Architecture

ManuSearch deconstructs the deep search process into three distinct agents:

1. Solution Planning Agent: This agent formulates sub-queries iteratively, serving as a strategic planner for decomposing complex questions. It operates using a ReAct-style architecture to manage the input context and generate questions based on previous reasoning steps.
2. Internet Search Agent: Deployed for real-time web searches, this agent retrieves pertinent documents to provide up-to-date evidence required for sub-queries. The search agent interacts dynamically, using tools such as web search engines and answer generation utilities to gather and process information.
3. Structured Webpage Reading Agent: After retrieving webpage data, this module cleans and processes raw HTML content to extract relevant information efficiently, thereby reducing the noise from web content. It performs this task using predefined intents to guide the extraction process.

Benchmarking and Results

The authors introduce ORION, a specialized benchmark focusing on reasoning over long-tail entities across various domains. ORION emphasizes tasks that require diverse cognitive abilities and information synthesis, thus presenting substantial challenges for current LLM systems.

ManuSearch's performance was evaluated on ORION alongside other datasets like FRAMES and GAIA. The experimental outcomes revealed that ManuSearch significantly surpasses previous open-source baselines, even outperforming leading closed-source systems on certain metrics. This underscores its potential as a high-fidelity alternative to proprietary models, promoting reproducible research and fostering innovation in open-source deep search technologies.

Implications and Future Directions

The paper highlights the critical need for transparent and extensible architectures in deep search systems. ManuSearch offers a foundation that allows researchers to examine, debug, and enhance each component, promoting trust and flexibility in system design. The framework's modularity is especially pertinent given the swift advancements in web-augmented LLMs and the pressure for systems to handle increasingly complex queries.

Potential future research directions include the integration of additional modalities, such as visual or auditory information retrieval, and the application of more sophisticated reinforcement learning techniques to further bolster reasoning capabilities in LLMs.

Overall, ManuSearch sets a new standard for transparent deep reasoning frameworks within the open-source community, emphasizing the importance of modular and interpretable system architectures in the field of AI.