Budget-Aware LLM Discovery via Cost-Calibrated Frontier Utility

The supplied abstract introduces CostAda, a controller that allocates inference-time discovery effort using quality progress, realized action cost, and remaining token budget. It claims that on 12 of 16 benchmark-backbone pairs, CostAda matches the strongest baseline’s full-budget quality with no more than half the budget, and reports the best mean final quality across eight benchmarks under GLM-5 and GPT-5.4. These results cannot be independently assessed from the abstract alone; the 2607 arXiv identifier, July 2026 publication date, and named model versions require verification.