Cost-Calibrated Frontier Utility for Budget-Aware LLM Discovery
A recent paper on arXiv (2607.26828) presents CostAda, an adaptive controller designed for the discovery of large language models (LLMs) that incorporates cost calibration. Traditional adaptive discovery controllers typically allocate credit based solely on score improvements, neglecting the varying token costs associated with different search actions, such as prompt length, retries, and guidance calls. The authors demonstrate that ignoring these costs can lead to a significant loss in achievable quality as the number of frontiers increases and costs vary. With a fixed token budget for searches, the controller must evaluate which frontier shows progress and whether the benefits outweigh the incurred costs before the budget runs out. CostAda employs a cost-calibrated utility for frontiers, assessing progress in relation to the costs of actions and adjusting credit based on the remaining budget, thereby influencing the intensity of local exploration. This paper appears to be a cross submission from a research group in computer science or AI.
Key facts
- Paper arXiv:2607.26828 introduces CostAda.
- CostAda is a cost-calibrated adaptive controller for LLM discovery.
- Existing controllers assign credit based only on score progress.
- Search actions incur different token costs (prompt length, retries, guidance calls).
- Cost-blind credit can forfeit all but a vanishing fraction of attainable quality.
- Controller must decide frontier improvement and justify cost under fixed token budget.
- CostAda uses cost-calibrated frontier utility.
- Utility conditions credit on remaining budget and controls local exploration intensity.
Entities
Institutions
- arXiv