Selecting the Optimal Prompter LLM Under Generalizability and Cost Constraints

Determine the optimal large language model to serve as the prompter in Booster’s guided recommendation pipeline that balances generalizability across environments (e.g., cross-schema transfer) and inference cost, identifying performance–cost trade-offs for different LLM scales and capabilities.

Background

The authors evaluate multiple embedder–prompter combinations and observe that larger LLMs can yield better configurations, especially for cross-schema scenarios, but at higher inference cost. They note the absence of a principled method for choosing an LLM given varying generalizability requirements and budget constraints.

This selection problem is left for future work, framing a concrete need to systematize how to choose or adapt LLMs for Booster across deployment contexts while controlling resource usage.

References

We leave the selection of the optimal LLM based on generalizability (e.g., cross-schema) and cost requirements for future work.

— This is Going to Sound Crazy, But What If We Used Large Language Models to Boost Automatic Database Tuning Algorithms By Leveraging Prior History? We Will Find Better Configurations More Quickly Than Retraining From Scratch!  (2510.17748 - Zhang et al., 20 Oct 2025) in Section 7.4 (Sensitivity Experiments: Embedder-Prompter Models)

This leaves open the question of whether optimization on high-capacity models would yield similar or superior prompts, and how these might transfer across model families.

— Larger Context Window, Fewer Overcorrections: Optimizing Prompts and Batching for Minimal-Edit Grammatical Error Correction  (2609.10810 - Karpo et al., 9 Sep 2026) in Section ‘Limitations’, item 3