InfoAgent: Advancing Autonomous Information-Seeking Agents

2025-09-29

AI Summary

This work addresses key limitations of existing agents—overreliance on black-box commercial search APIs, weak tool interaction capabilities, and poor scalability—by proposing InfoAgent, an open-source agent framework supporting autonomous web search. Methodologically, it introduces a self-hosted search infrastructure and a problem construction strategy based on entity trees and subtree sampling; designs an entity obfuscation pipeline for synthetic data generation; and employs a two-stage post-training paradigm (supervised fine-tuning followed by reinforcement learning) to enhance multi-step reasoning and tool invocation. Evaluated on BrowseComp, BrowseComp-ZH, and Xbench-DS, InfoAgent achieves accuracy scores of 15.3%, 29.2%, and 40.4%, respectively—substantially outperforming strong baselines such as WebSailor-72B. These results demonstrate InfoAgent’s effectiveness and reproducibility in long-horizon, reasoning-driven information retrieval tasks.

Technology Category

• Search and Optimization: Learning to Search
• Natural Language Processing: Information Extraction
• Multiagent Systems: Modeling other Agents

Application Category

• Search and Retrieval-Augmented AI: Agentic search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Building Large Language Model agents that expand their capabilities by interacting with external tools represents a new frontier in AI research and applications. In this paper, we introduce InfoAgent, a deep research agent powered by an innovative data synthesis pipeline and orchestrated web search tools. To construct challenging, hard-to-find queries,we build entity trees and apply sub-tree sampling with entity fuzzification to systematically increase question difficulty. Unlike prior work that relies heavily on commercial search tools, we develop a dedicated self-hosted search infrastructure, enhancing transparency of agent environments and facilitating further advancement of agent capacity. We evaluate the effectiveness of our data pipeline by measuring the average number of tool calls required to correctly answer a question, and also show that our agent yields better performance when equipped with our tools. Our mbox{InfoAgent} is post-trained from Qwen3-14B using a two-stage recipe: cold-start supervised finetuning to instill long-horizon search behaviors, followed by reinforcement learning which significantly improves reasoning-driven tool use. With our methods, InfoAgent achieves 15.3% accuracy on BrowseComp, 29.2% on BrowseComp-ZH, and 40.4% on Xbench-DS, outperforming prior open-source deep research agents such as WebSailor-72B and DeepDive-32B.

Problem

• Building autonomous agents that expand capabilities through external tools
• Developing transparent self-hosted search infrastructure for agent environments
• Enhancing reasoning-driven tool use via staged training methodology

Innovation

• Entity tree sampling with fuzzification for query generation
• Self-hosted search infrastructure for transparent agent environments
• Two-stage post-training combining supervised finetuning and reinforcement learning