SciResearcher: Scaling Deep Research Agents for Frontier Scientific Reasoning

2026-05-02

AI Summary

This work addresses the challenge that cutting-edge scientific research relies on fragmented and heterogeneous academic resources, which existing AI agents struggle to leverage for complex reasoning beyond simple fact retrieval. To bridge this gap, we propose SciResearcher, an automated agent framework that introduces a novel synthetic data generation paradigm tailored for scientific reasoning. Our approach integrates literature mining, task synthesis, tool invocation, supervised fine-tuning, and agent-based reinforcement learning to enhance capabilities in information acquisition, tool integration, and long-horizon planning. The resulting SciResearcher-8B model achieves 19.46% on the HLE-Bio/Chem-Gold benchmark and demonstrates absolute performance gains of 13–15% on SuperGPQA-Hard-Biology and TRQA-Literature, setting new state-of-the-art results among models of comparable scale.

Abstract

Frontier scientific reasoning is rapidly emerging as a key foundation for advancing AI agents in automated scientific discovery. Deep research agents offer a promising approach to this challenge. These models develop robust problem-solving capabilities through post-training on information-seeking tasks, which are typically curated via knowledge graph construction or iterative web browsing. However, these strategies face inherent limitations in frontier science, where domain-specific knowledge is scattered across sparse and heterogeneous academic sources, and problem solving requires sophisticated computation and reasoning far beyond factual recall. To bridge this gap, we introduce SciResearcher, a fully automated agentic framework for frontier-science data construction. SciResearcher synthesizes diverse conceptual and computational tasks grounded in academic evidence, while eliciting information acquisition, tool-integrated reasoning, and long-horizon capabilities. Leveraging the curated data for supervised fine-tuning and agentic reinforcement learning, we develop SciResearcher-8B, an agent foundation model that achieves 19.46% on the HLE-Bio/Chem-Gold benchmark, establishing a new state of the art at its parameter scale and surpassing several larger proprietary agents. It further achieves 13-15% absolute gains on SuperGPQA-Hard-Biology and TRQA-Literature benchmarks. Overall, SciResearcher introduces a new paradigm for automated data construction for frontier scientific reasoning and offers a scalable path toward future scientific agents.

Problem

• frontier scientific reasoning
• deep research agents
• heterogeneous academic sources
• complex computation
• scientific discovery

Innovation

• frontier scientific reasoning
• automated data construction
• agentic framework
• tool-integrated reasoning
• scientific discovery