Hugging Face Overview
Overview
FractalAIResearch's Fathom-Search-4B is a 4-billion parameter model designed for advanced information retrieval and synthesis from live web content. It is a key component of the Fathom-DeepResearch system, which has achieved state-of-the-art performance in the open-weights category on search-intensive benchmarks like SimpleQA, FRAMES, WebWalkerQA, and Seal0. The system also outperforms several closed-source DeepResearch agents on the DeepResearch-Bench for open-ended synthesis.
Key Capabilities
- Live Web Search: Optimized for long-horizon, evidence-seeking through real-time web browsing.
- Information Extraction & Verification: Capable of extracting and verifying information from diverse web sources.
- Agentic System: Works in conjunction with Fathom-Synthesizer-4B to form a comprehensive DeepResearch agent.
- Robust Search Backend: Utilizes a specialized search tool server built on Jina-AI, Crawl4AI, Trafilatura, and Serper.dev, supporting various sources including YouTube, PDFs, Reddit, and GitHub.
Key Innovations
- Multi-Agent Self-Play: A self-supervised dataset construction framework for generating verifiable, live web-search enforcing, multi-hop QA pairs (DUETQA dataset).
- RAPO (Reward-Aware Policy Optimization): A zero-overhead extension of GRPO for stabilizing multi-turn Reinforcement Learning with Verifiable Rewards.
- Steerable Step-Level Reward: A novel reward function to alleviate reward-hacking in RLVR training, allowing steering of tool usage and cognitive allocation.
- DeepResearch Report Synthesis Protocol: A plan-then-write protocol for synthesizing DeepSearch traces into citation-dense reports, involving question decomposition, evidence-to-section mapping, and insight planning.
Good For
- Applications requiring extensive and verifiable information retrieval from the live web.
- Developing agentic systems that need to reason over dynamic, real-time data.
- Generating detailed, citation-dense reports based on deep web investigations.
- Research and development in long-horizon information retrieval and synthesis for Small Language Models (SLMs).