CostAda Optimizes LLM Discovery Under Fixed Token Budgets

This research introduces CostAda, an adaptive controller designed to enhance the efficiency of large language model (LLM) discovery processes, particularly when operating under strict token budgets. Current discovery controllers often fail to account for the varying token costs associated with different search actions, such as prompt length or retry attempts, leading to inefficient resource allocation. CostAda addresses this by implementing a "cost-calibrated frontier utility" system. CostAda evaluates the progress of a search frontier relative to its incurred cost and the remaining budget, using this signal to dynamically control exploration intensity, allocate resources across frontiers, and intervene with budgeted tactics. This approach ensures that the search process is shaped by cost and budget considerations from the outset, rather than merely using them as stopping rules. Empirical results show that CostAda can achieve the quality of the strongest baselines with up to half the token budget on many benchmarks, and it consistently delivers superior final quality across various LLMs.