Cited but Not Verified: Deep Research Agents Show Only 39-77% Factual Accuracy Despite 94% Valid Links

arXiv·high signal

Evaluation of LLM deep research agents reveals a critical gap: link validity exceeds 94% and relevance above 80%, but factual accuracy against source material is only 39-77%. When tool calls escalate from 2 to 150, fact-check accuracy drops ~42% on average — more research depth paradoxically reduces citation reliability. Fewer than half of open-source models produce properly cited reports.