LLMs can find better hyperparameters faster than random search in low-budget settings, speeding model iteration and cutting compute cost when trials are expensive.

PREFER automates prompt generation and ensemble weighting, yielding better few-shot accuracy and much lower API/time cost than heavy prompt-search methods.

You can cut prompt-evaluation costs and improve per-query outputs by training a small offline reward model on past prompt logs and using it to pick prompts instead of repeatedly calling expensive LLM verification.

Train a single lightweight retriever once (with a small model) to boost many larger LLMs at inference, cutting repeated fine-tuning costs and improving zero-shot accuracy on many NLU tasks.

SEE finds stronger prompts with far fewer API calls and tokens, so teams can improve LLM task accuracy while cutting prompt optimization cost and speeding experimentation.

HMAW automates prompt tuning without training and boosts response quality across varied tasks, letting teams improve outputs quickly while avoiding dataset-specific finetuning.

You can improve black-box LLM few-shot extraction by choosing demos that match the model's own output style; this yields production-quality gains without fine-tuning and reduces annotation cost for Open IE.

Automates and concentrates prompt tuning across complex LLM pipelines, reducing manual engineering time and often improving accuracy while lowering token costs.

SynthesizeMe yields interpretable persona prompts that improve in-context judgment of user preferences without full model finetuning; useful when collecting a few pairwise judgments is feasible but large-scale retraining is not.

You can improve the performance of an API-only LLM on a task without model access or costly fine-tuning. That lowers operational friction for domain teams who need better outputs fast.

SOPL finds better human-readable prompts with far fewer costly LLM evaluations by modeling prompt features and choosing experiments adaptively, lowering API costs and time for deploying LLM-based features.

AutoHD improves LLM planning accuracy without extra model training and produces interpretable Python heuristics you can inspect and reuse.

PolyPrompt offers a low-cost way to raise non-English QA accuracy by training a few small embeddings per language instead of costly model fine-tuning.

PMPO cuts evaluation cost by using log‑probabilities instead of sampling and external judges, enabling faster prompt tuning for deployed models and improving midsize model outputs without fine‑tuning.

meta-learning optimizer prompts and compositions can boost task accuracy and reduce model cost by letting cheaper program models be amplified by smarter optimizer/meta calls.

If you use LLM search or prompt optimization to improve accuracy, relying on standard generation-confidence scores risks optimizing for diverse but wrong outputs; invest in correctness-specific checks or new uncertainty estimators.

HbBoPs reduces the number of expensive LLM API calls needed to find a good static prompt, cutting cost and time for model-driven features that rely on single-prompt deployments.

GLaPE lets teams optimize prompts without costly labels, enabling prompt tuning for private models and new tasks while cutting annotation costs.

Automating algorithm discovery with LLMs can find improved model-compression strategies and reduce expert hours needed to hand-craft search heuristics.

Fixed-question testing can hide or undercount model knowledge and lead to poor model choices; optimized, semantics-preserving prompt search reveals a model's true answerable range so teams can pick models that actually cover needed domain facts.

SPT can raise task accuracy significantly without costly model fine-tuning; it offers a lower-barrier way to boost product QA and reduce hallucinations if you can afford extra API/computation and representative training data.

DFSL gives near state-of-the-art KGQA without dataset fine-tuning, cutting training cost and enabling faster deployment across knowledge graphs.

Automates prompt tuning for black-box LLMs, cutting manual time and raising held-out performance by ~14% on tested tasks.

Label fewer examples and get nearly the same or better in-context performance while cutting selection time and inference cost; this lowers annotation bills and speeds up prompt curation.