Our empirical findings uncover critical misconceptions in agentic research workflows:
•Longer Deliberation ≠ Better Retrieval: Explicit reasoning (THINK mode) yields no consistent gains; in Wide tasks, it often triggers redundant searches and latency without reducing hypothesis uncertainty.
•Full-Text Repositories Outperform Web Search: Replacing DeepXiv with open-web search causes Deep accuracy to drop from 5.42% to 3.97%, proving that web-visible snippets fail to capture paper-internal evidence.
•Divergent Test-Time Scaling: Scaling inference compute substantially boosts Deep Research by recovering brittle trajectory missteps, yet provides marginal gains on Wide tasks due to fundamental recall bottlenecks.