Balancing performance and efficiency in long-horizon agentic search

Determine effective strategies to balance long-horizon agentic search performance and computational efficiency for tool-augmented large language model research agents that conduct multi-step web search, browsing, and evidence aggregation under constrained interaction budgets and latency requirements.

Background

The paper observes that many deep research agents improve results by increasing reasoning depth and the number of tool calls, which in turn drives up inference latency and computational cost in search-intensive scenarios. This tension is especially acute for long-horizon tasks that require extensive external evidence acquisition and verification.

While the authors propose the SMTL framework to address efficiency via parallel evidence acquisition and structured context management, they explicitly note that the broader challenge of balancing task performance with compute and latency constraints for long-horizon agentic search remains unsettled.