TLDR: InfoAgent is a new deep research agent powered by Large Language Models (LLMs) that excels at complex information-seeking tasks. It uses an innovative data synthesis method to create challenging questions and a custom web search infrastructure for efficient training. InfoAgent, based on Qwen3-14B, is trained in two stages (supervised finetuning and reinforcement learning) and achieves state-of-the-art performance among open-source models, even outperforming some larger ones on various benchmarks.
In the rapidly evolving landscape of artificial intelligence, the development of Large Language Model (LLM) agents that can autonomously interact with external tools is a significant frontier. A new research paper introduces InfoAgent, a sophisticated deep research agent designed to push the boundaries of autonomous information-seeking.
Authored by a collaborative team from Southeast University, Brown University, and Microsoft, InfoAgent stands out by leveraging an innovative data synthesis pipeline and a custom-built web search infrastructure. This agent aims to tackle complex, multi-step information-seeking problems that require deep reasoning and extensive tool use.
The Challenge of Deep Research Agents
Current deep research agents (DRAs) often face two primary hurdles: generating sufficiently challenging training data and establishing an efficient, stable interactive environment. Many existing open-source DRAs are trained on relatively simple datasets, leading to shallow search behaviors. Furthermore, reliance on commercial search APIs can introduce issues with transparency, cost, and consistency during training.
InfoAgent’s Innovative Approach
InfoAgent addresses these challenges head-on with two key innovations:
1. Advanced Data Synthesis Pipeline: To create truly challenging questions, InfoAgent employs a two-stage data synthesis pipeline. It starts by constructing ‘entity trees’ from Wikipedia entities. Then, it uses ‘sub-tree sampling’ combined with ‘entity fuzzification.’ This process systematically increases question difficulty by replacing specific names and dates with generic descriptions or ranges, forcing the agent to perform long-horizon retrieval and complex reasoning rather than simple lookups.
2. Dedicated Self-Hosted Search Infrastructure: Unlike many prior works that depend on commercial search tools, InfoAgent utilizes its own self-hosted search and browsing infrastructure. This provides fine-grained control over outputs, enhances transparency, and ensures high throughput and consistency, which are crucial for stable reinforcement learning training.
How InfoAgent Works
InfoAgent operates within the ReAct framework, iteratively combining reasoning with tool calls. It uses two primary web-based tools:
- Search Function: Given a query, it retrieves a ranked list of web URLs along with concise snippets. To improve snippet quality, it crawls full web pages, splits them into chunks, and uses a combination of BM25, embedding models, reranker models, and an LLM to generate highly relevant and compact summaries.
- Browse Function: Given a URL, it returns a detailed semantic document of the corresponding web page, allowing for deeper investigation. This mimics human behavior when exploring interesting search results.
Training and Performance
InfoAgent is post-trained from Qwen3-14B using a two-stage recipe:
- Cold-Start Supervised Finetuning (SFT): This initial stage instills long-horizon search behaviors into the model.
- Reinforcement Learning (RL): Following SFT, reinforcement learning significantly refines the agent’s ability for reasoning-driven tool use.
The results are impressive. InfoAgent achieves state-of-the-art performance among open-source models with fewer than 15 billion parameters. For instance, it scores 15.3% accuracy on BrowseComp, 29.2% on BrowseComp-ZH, and 40.4% on Xbench-DS. Notably, it even surpasses some much larger open-source models, such as WebSailor-72B and DeepDive-32B, on several benchmarks. Despite being trained exclusively on English data, InfoAgent demonstrates strong cross-lingual generalization on Chinese benchmarks.
Ablation Studies and Insights
The research also includes ablation studies that highlight the importance of the training components:
- The SFT cold start is crucial for teaching complex agent tasks like deep research, as base models often lack these abilities.
- The quality of the custom search tool significantly impacts performance, providing more useful information for tackling difficult problems compared to simpler retrievers.
- Training on longer trajectories (more tool calls) generally leads to higher accuracy on challenging tasks, though a balance is needed to prevent excessive tool calling.
Also Read:
- Deepening AI’s Understanding: The Fathom-DeepResearch System Explained
- KGQAGen: A Framework for High-Quality Knowledge Graph Question Answering Datasets
Looking Ahead
While InfoAgent marks a significant step forward, the authors acknowledge limitations, such as the current context length restriction during reinforcement learning and the reliance on the Wikipedia corpus for data synthesis. Future work aims to address these by adopting models with longer context windows and expanding the data pipeline to the broader web.
InfoAgent represents a compelling advancement in building autonomous information-seeking agents, demonstrating how innovative data synthesis and custom tool design can unlock powerful deep research capabilities in LLMs. You can read the full paper here: InfoAgent: Advancing Autonomous Information-Seeking Agents.
