TLDR: InfoSeek is a novel, open-source framework that synthesizes large-scale, complex Deep Research tasks, formalizing them as Hierarchical Constraint Satisfaction Problems (HCSPs). It uses a dual-agent system to build ‘Research Trees’ from web data, generating verifiable, multi-step reasoning questions. Models like InfoSeeker-3B, trained on this dataset, demonstrate superior performance on challenging benchmarks, outperforming larger models and commercial APIs, and advancing LLMs’ capabilities in complex information seeking and reasoning.
Large language models (LLMs) are becoming increasingly sophisticated, moving beyond simple factual questions to tackle what researchers call “Deep Research.” This involves breaking down complex problems, coordinating multi-step reasoning, and bringing together information from many different sources. However, current methods and datasets haven’t fully captured the complexity of these tasks, often leading to simplified reasoning or issues like knowledge leakage.
To bridge this gap, a new framework called InfoSeek has been introduced. InfoSeek is designed to create complex Deep Research tasks in a scalable way. It formalizes these tasks as Hierarchical Constraint Satisfaction Problems (HCSPs), which are fundamentally different from simpler problem types like single-constraint or multi-hop questions. HCSPs require solving a hierarchy of interdependent sub-problems to arrive at a unique and verifiable answer.
How InfoSeek Works
InfoSeek uses a clever dual-agent system to build what’s called a “Research Tree.” This tree is constructed recursively from vast amounts of web pages and Wikipedia data. The system blurs intermediate nodes into valid sub-problems and then converts these structured trees into natural language questions that demand traversing the entire hierarchy for a solution. This ensures that the generated questions require genuine multi-step reasoning and yield unique, verifiable answers.
The framework involves two main agents: a Planner and a Browser. The Planner oversees the construction of the Research Tree, deciding where to add new information or constraints. The Browser then executes these plans by searching webpages, extracting relevant information, and validating it. This process allows for the systematic creation of complex problems with controlled difficulty.
A crucial aspect of InfoSeek is its data quality assurance. The framework ensures that the generated problems are challenging enough that even powerful LLMs cannot solve them using only their existing knowledge. It also verifies that each question is factually grounded and solvable through the generated search trajectory, preventing ambiguity or unsolvable problems.
The InfoSeek Dataset and InfoSeeker Model
InfoSeek has successfully generated a large-scale dataset comprising over 50,000 training examples and 16,500 reasoning trajectories. This dataset is fully open-source, encouraging further research and development in the community. The complexity of the dataset is evident in the high failure rates of even powerful baseline models, confirming its suitability for training advanced research agents.
Building on this dataset, the researchers developed InfoSeeker, a framework for advanced agentic search and deep research. InfoSeeker employs a novel workflow that includes parallel multi-query search and a dedicated Refiner Agent. This Refiner Agent condenses retrieved information into concise summaries, maintaining high recall while keeping the working context manageable for the LLM. The model also incorporates a “Think Before Action” phase, encouraging reflection and more targeted query generation.
InfoSeeker is trained using a two-stage process: supervised fine-tuning (SFT) with rejection sampling to learn from successful reasoning trajectories, followed by reinforcement learning (RL) to further enhance its reasoning and search capabilities. This pipeline enables even compact LLMs to achieve impressive performance.
Also Read:
- DeepResearch Arena: A New Benchmark to Test AI’s Research Acumen
- AnchorRAG: A Multi-Agent Framework for Enhanced Open-World Question Answering with Knowledge Graphs
Impressive Performance
Experiments show that InfoSeeker-3B, a relatively small LLM, consistently outperforms strong baselines on both single-hop and multi-hop question answering benchmarks. More notably, on the challenging BrowseComp-Plus benchmark, which focuses on complex, search-intensive reasoning, InfoSeeker-3B surpasses much larger 32B models and lightweight commercial APIs like Gemini 2.5 Flash and Sonnet 4. It even achieves performance comparable to stronger APIs such as Gemini 2.5 Pro.
This remarkable performance highlights the efficiency of the InfoSeek framework in scaling down deep research capabilities to more compact LLMs. The research also demonstrates that training on the purpose-built InfoSeek dataset yields substantially stronger deep research performance and more effective use of search tools compared to traditional datasets like Natural Questions and HotpotQA.
The development of InfoSeek and InfoSeeker marks a significant step forward in enabling LLMs to perform complex Deep Research tasks. By providing a scalable, high-quality dataset and an effective training framework, this work paves the way for the next generation of LLMs to become autonomous knowledge engines. You can find more details in the full technical report: OPEN DATA SYNTHESIS FOR DEEP RESEARCH.
