New Deep Research Pretraining framework enhances AI agent training

Researchers have developed Deep Research Pretraining (DRP), an offline framework designed to improve the training of deep research agents. DRP derives supervision from existing evidence structures like citation graphs and hyperlinks, converting them into search-open-write trajectories. This method teaches models to effectively search, inspect documents, and synthesize evidence without requiring a live retrieval environment. When tested on scholarly citation graphs (DRP-Paper) and Wikipedia hyperlinks (DRP-Web), DRP consistently outperformed models trained without it, even achieving superior results with less data and showing benefits in downstream agentic reinforcement learning. AI

IMPACT This framework could significantly reduce the cost and complexity of training advanced AI research agents, potentially accelerating their development and deployment in complex information-seeking tasks.

RANK_REASON The cluster describes a new research paper detailing a novel pretraining framework for AI agents. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• DeepResearch Bench
• Deep Research Pretraining
• DRP-Paper
• DRP-Web
• Qwen3-14B-Base
• ResearchQA
• SimpleQA
• WebWalkerQA

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →