TLDR: Fathom-DeepResearch is an AI system designed to improve how language models find and synthesize information from the web, especially for complex, open-ended questions. It consists of two main models: Fathom-Search-4B, which uses advanced techniques like the DUETQA dataset, RAPO, and a steerable reward system to perform extensive web searches and tool interactions; and Fathom-Synthesizer-4B, which then transforms these search results into detailed, structured reports. The system achieves state-of-the-art performance in open-source categories and rivals proprietary models, enabling AI agents to conduct more thorough and accurate investigations.
In the rapidly evolving landscape of artificial intelligence, a significant challenge for Large Language Models (LLMs) has been their ability to perform complex, open-ended information-seeking tasks that require extensive web interaction and sophisticated synthesis. While proprietary AI systems have shown remarkable capabilities in this area, open-source alternatives often struggle with sustained tool usage and comprehensive information synthesis. Addressing this gap, researchers from Fractal AI Research have introduced Fathom-DeepResearch, an innovative agentic system designed to unlock long-horizon information retrieval and synthesis for LLMs.
The Fathom-DeepResearch system is built upon two specialized models: Fathom-Search-4B and Fathom-Synthesizer-4B. Together, these models aim to elevate the performance of open-source AI agents in deep information exploration and report generation.
Fathom-Search-4B: The Web Investigator
Fathom-Search-4B is the core component responsible for evidence-based investigation through live web search and targeted webpage querying. It’s trained from the Qwen3-4B model and incorporates three key advancements to enhance its capabilities:
- DUETQA Dataset: This is a unique dataset of approximately 5,000 samples, generated through a multi-agent self-play process. What makes DUETQA special is that its questions are specifically designed to be unanswerable without live web search, ensuring that models learn to depend on external tools for information. It also enforces the use of diverse and heterogeneous web sources.
- RAPO (Reward-Aware Policy Optimization): A zero-overhead extension of GRPO, RAPO is crucial for stabilizing multi-turn Reinforcement Learning (RL) with verifiable rewards. It tackles common issues like training instability and reward hacking by using curriculum pruning, reward-aware advantage scaling, and per-prompt replay buffers. This allows Fathom-Search-4B to reliably extend its tool-calling abilities beyond 20 calls, which is vital for deep information exploration.
- Steerable Step-Level Reward: This innovative reward system classifies each tool call by its cognitive behavior and marginal utility. It provides explicit control over the breadth, depth, and horizon of the search trajectory, preventing the agent from making redundant tool calls and encouraging efficient, diverse information seeking.
These improvements enable Fathom-Search-4B to perform sustained and efficient web interactions, a critical requirement for tackling high-uncertainty reasoning tasks.
Fathom-Synthesizer-4B: The Report Generator
Complementing the search capabilities, Fathom-Synthesizer-4B, also trained from Qwen3-4B, is designed to convert the multi-turn DeepSearch traces generated by Fathom-Search-4B into structured, citation-dense DeepResearch Reports. This model follows a unique “Plan-then-Write” protocol, where it first decomposes the query into sub-goals, defines the report structure, maps evidence to specific sections, and outlines strategies for generating insights. Only after this planning phase does it produce the final report, complete with citations strictly drawn from the URLs explored by Fathom-Search-4B.
The training of Fathom-Synthesizer-4B utilizes the DEEPRESEARCH-SFT dataset, a synthetic corpus distilled from GPT-5. This dataset provides supervision for question decomposition, section mapping of evidence, and planning for insights, ensuring high-quality, well-structured, and accurately cited reports.
Also Read:
- G-reasoner: Unifying Graph and Language Models for Advanced Knowledge Reasoning
- Unraveling How Large Reasoning Models Arrive at Answers
Performance and Impact
Fathom-DeepResearch has been rigorously evaluated on various DeepSearch benchmarks, including SimpleQA, FRAMES, WebWalker, Seal0, and MuSiQue, as well as the specialized DeepResearch-Bench. The system has achieved state-of-the-art performance in the open-weights category, demonstrating strong generalization across diverse reasoning tasks such as HLE, AIME-25, GPQA-Diamond, and MedQA. Notably, it consistently outperforms its base model and other open-source systems, and even surpasses larger closed-source models like GPT-4o in many challenging DeepSearch tasks.
On open-ended benchmarks like DeepResearch-Bench, Fathom-DeepResearch rivals and often outperforms many proprietary closed-source systems, including Claude, Grok, and Perplexity Deep Research. This underscores its competitiveness in end-to-end deep research tasks and its ability to generate comprehensive, factual, and well-cited reports.
The introduction of Fathom-DeepResearch marks a significant step forward for open-source AI, providing a robust and capable system for complex information retrieval and synthesis. For more technical details, you can refer to the original research paper.
