テクニカルガイド

Automatic Prompt Optimization and DSPy

Automatic prompt optimization uses an algorithm to search for better instructions or few-shot examples.

4 分で読めます
最終更新日 2026年9月15日

このページでは4 分で読めます

概要
ディープダイブ
戦略的影響
The Future of Automatic Prompt Optimization and DSPy
現実世界の実装
リスクとガードレール
実装ロードマップ
探検を続けましょう
よくある質問

概要

It scores candidate prompts against a metric on a set of examples instead of relying on hand-tweaked wording. Well-known approaches include APE, Google DeepMind's OPRO and DSPy, an open-source framework from Stanford. Together they turn prompt engineering from guesswork into a measurable, repeatable process.

ディープダイブ

Hand-tuned prompts have a known weakness. Wording that works for one model, dataset or version can quietly stop working on another, and changes are usually judged on only a handful of examples. Automatic prompt optimization replaces that loop with a search. You provide three things: - A task. - A set of inputs, ideally with expected outputs. - A metric, such as exact match or a rubric scored by another model. The optimizer generates candidate prompts, runs them, scores them and keeps the winners. Early examples made the idea concrete. APE (Automatic Prompt Engineer, Zhou et al., 2022) had a model propose many instructions from input-output examples, then kept the highest-scoring ones. OPRO (Optimization by PROmpting, Yang et al., Google DeepMind, 2023) used an LLM as the optimizer. Its meta-prompt listed earlier instructions with their scores, and the model proposed new ones. OPRO found instructions such as "Take a deep breath and work on this problem step-by-step," which scored well for one particular model on grade-school math. Optimized wording can be unintuitive and specific to a model. DSPy goes further by changing how programs are written. It was developed at Stanford NLP by Omar Khattab and collaborators and introduced in 2023. You declare what each step does with a signature, such as "question, context -> answer," and combine modules like Predict or ChainOfThought. An optimizer (earlier called a teleprompter), such as BootstrapFewShot or MIPROv2, then "compiles" the program by choosing demonstrations and instructions for each module to maximize your metric. Three misconceptions are worth correcting: - These tools remove human judgment. They do not, because the metric and examples define "better." - Optimized prompts are safe to trust. They can overfit a small dataset. - Optimization is free. It costs real API calls, sometimes many.

技術的な洞察

In DSPy, what gets optimized is each module's prompt, which has two tunable parts: the instruction text and the demonstrations. BootstrapFewShot runs your program on the training inputs and keeps the traces where the metric passes. It uses those traces' intermediate inputs and outputs as few-shot examples. This means you can get demonstrations for internal steps you never labeled. MIPROv2 also proposes candidate instructions. It then searches over combinations of instructions and demonstrations, using Bayesian optimization with a surrogate model to decide which combinations to try. In practice you need three things: a metric function that returns a score or a boolean, a training set, and a separate held-out set to check for overfitting. Budget matters, because each trial is a batch of model calls. A common pattern is: 1. Optimize with a cheaper model when possible. 2. Verify on the target model. 3. Save the compiled program. 4. Rerun the optimizer whenever the model or the data changes.

戦略的影響

費用と予算

アーキテクチャの決定により、パフォーマンスと運用コストが何年にもわたって推進されます。

より明確な判決

技術教育は、チームが最新のスタックだけでなく、適切なスタックを選択するのに役立ちます。

品質管理

より良いエンジニアリングの選択により、本番環境での信頼性に関するインシデントが減少します。

The Future of Automatic Prompt Optimization and DSPy

Prompt optimization is becoming a normal part of LLM engineering, alongside evaluation suites, rather than a research curiosity. Open questions remain. How well do optimized prompts transfer between models? How do you write metrics for subjective qualities like tone or helpfulness? How do you avoid overfitting small datasets? Some research combines prompt optimization with lightweight fine-tuning, and DSPy supports both. Models change often, so the ability to re-optimize automatically against a stable test set is likely to matter more than any single clever prompt. Defining good metrics and examples remains a human job.

現実世界の実装

A team building a support-ticket classifier writes a DSPy signature "ticket -> category" and supplies 200 labeled tickets and an accuracy metric. An optimizer then picks the few-shot demonstrations that score best.

When a company moves from one LLM provider to another, it reruns its DSPy optimizer for the new model. It does not hand-rewrite prompts that were tuned for the old model.

A researcher runs an OPRO-style loop. One model proposes new wordings for an instruction, each wording is scored on a math benchmark, and the best-scoring history goes back into the next round of proposals.

A retrieval-augmented question-answering pipeline is optimized end to end, with a metric that checks final answers against reference answers. Both the query-writing step and the answering step improve together.

リスクとガードレール

1 つのベンチマークを最適化すると、より広範なシステムの弱点が隠れる可能性があります。
インフラストラクチャとメンテナンスのコストは過小評価されがちです。
システムが複雑になるにつれて、セキュリティと可観測性のギャップが拡大する可能性があります。

実装ロードマップ

実装前にレイテンシ、品質、コストの目標を定義します。
現実的な負荷とデータ条件でのベンチマーク。
エラー、ドリフト、ユーザーへの影響を計測器で監視します。
スケーリングの前に、ロールバックとインシデント対応のパスを準備します。

探検を続けましょう

AI ベンチマーク

技術的なオプションを比較するときは、評価を適切に使用してください。ガイドを読む

強化学習

技術的なトレーニング戦略をさらに深く掘り下げます。ガイドを読む

Free newsletter

Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.

One email each weekday. Unsubscribe in one click. We never sell or share your address.

Test yourself

Take the Automatic Prompt Optimization and DSPy quiz

Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.

クイズを開始する

Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation

よくある質問

What is Automatic Prompt Optimization and DSPy?

Automatic prompt optimization uses an algorithm to search for better instructions or few-shot examples. It scores candidate prompts against a metric on a set of examples instead of relying on hand-tweaked wording. Well-known approaches include APE, Google DeepMind's OPRO and DSPy, an open-source framework from Stanford. Together they turn prompt engineering from guesswork into a measurable, repeatable process.

What must you supply for automatic prompt optimization to work?

The optimizer needs a task, examples and a metric so it can score candidate prompts and keep the best ones.

In OPRO, what plays the role of the optimizer?

OPRO uses a language model as the optimizer. It sees past instructions with their scores and proposes new ones.

What does OPRO's "Take a deep breath and work on this problem step-by-step" result illustrate?

The instruction scored well for one particular model on grade-school math. Wording found by search can be surprising and may not carry over to other models.

What is a DSPy signature?

Signatures declare what a module should do in terms of its input and output fields. DSPy then works out the prompt.

How does DSPy's BootstrapFewShot get its demonstrations?

Successful traces become few-shot examples, including for internal steps you never labeled yourself.

学び続ける