Deep Research Agents
Overview
Deep research agents use search and other tools to investigate questions that require gathering evidence across many sources, reasoning over it, and producing structured reports. Common approaches include ReAct-style search, query refinement and reranking, explicit plans or evidence graphs, and multi-agent delegation; systems are trained with supervised trajectories and reinforcement learning, sometimes guided by rubrics. Evaluation spans web-search benchmarks, expert-scored reports, and process-level checks of evidence, citations, and intermediate decisions. Agents can produce convincing reports, but completeness, resistance to misleading or conflicting sources, alignment with user intent, and the cost of long investigations remain difficult to balance.
Latest papers
115 papers18 papers in the last four weeks, up 100% on the four weeks before. 0.1% of all new papers.