Budget Control for LLM Search Agents via Value-of-Information

A recent preprint on arXiv (2605.05701) presents a novel technique for managing inference-time budgets in search agents utilizing LLMs. This method tackles the issue of simultaneous constraints on both tool usage and token generation. The authors conceptualize multi-hop question answering as a two-phase budget management challenge. During the search phase, a controller evaluates each viable action by assigning a Value-of-Information (VOI) score, which estimates the additional task value per budget unit based on the current state and remaining dual budget. This scoring system informs choices regarding retrieval, decomposition, and answer commitment. Following the search, a selective evidence-based finalizer assesses the trajectory answer against a refined candidate. The objective is to enhance answer quality under strict limitations without the need for more advanced models.