inference-time-control

This paper introduces a two-stage inference-time budget control method for LLM search agents, using Value-of-Information scores to optimize tool-call and token allocation during multi-hop question answering.