Cost-Aware Cross-Tier Transfer Reduces LLM Prompt Optimization Costs

A new study published on arXiv (2608.10694v2) introduces an innovative and budget-friendly approach for enhancing LLM prompts and agent programs using evolutionary methods. In this technique, the LLM's functions are divided during the search: a low-cost tier manages high-volume outputs, while a stronger model takes on rare tasks requiring deeper reflection. The approach leverages upward cross-tier transfer to apply a cost-effective prompt to a more advanced model. The research details a framework for deciding when cheaper searches can substitute for more expensive ones, based on testing across four tasks (HotpotQA, IFBench, LiveBench-Math, HoVer) with eleven models. Results show a performance equivalent or superior to same-tier optimizations, achieving significant cost reductions of 5.6-14x, and even 25-54x in certain cases.