EvoPrompt (EvoPrompt)

What is EvoPrompt?

EvoPrompt is a research technique introduced by Guo et al. in 2023 that applies evolutionary computation to the problem of automatic prompt optimization. Instead of having a human iterate on prompt variants manually, EvoPrompt maintains a population of candidate prompts and uses evolutionary operators — selection, crossover, and mutation — to search the space of possible prompts guided by objective task performance.

The critical insight is that the crossover and mutation operations are themselves performed by an LLM: rather than splicing text randomly (as in traditional genetic algorithms), EvoPrompt prompts an LLM to intelligently combine or modify candidate prompts in semantically meaningful ways. This makes the search space exploration far more efficient than random perturbation.

The algorithm operates in three phases, repeated for N generations:

• Evaluation (Fitness): Each candidate prompt in the population is run against a labeled development set. Task performance (accuracy, F1, etc.) becomes the fitness score.
• Selection: The highest-fitness candidates are selected as parents for the next generation. Poor performers are discarded.
• Variation (Crossover + Mutation): An LLM combines pairs of selected prompts (crossover) and makes targeted modifications to individual prompts (mutation) to produce new offspring candidates for the next generation.

When to Use EvoPrompt

How to Implement EvoPrompt

1. 1

   Define the task and fitness function

   Specify what task you are optimizing for and how to measure prompt performance. For classification, this is typically accuracy on a held-out development set of 50–200 labeled examples. The fitness function must be computable automatically — EvoPrompt runs dozens of evaluation cycles and cannot rely on human judgment for each.

2. 2

   Initialize a diverse population of seed prompts

   Create 10–20 candidate prompts with significant structural and stylistic diversity. Include: direct task instructions, persona-based prompts, few-shot example prompts, chain-of-thought prompts, and format-specific prompts. Diversity in the initial population is essential — EvoPrompt cannot explore what is not represented in the gene pool.

3. 3

   Evaluate fitness for each candidate

   Run every candidate prompt on all development set examples and compute the fitness score. This is computationally expensive — with 20 candidates and 100 dev examples, that is 2,000 LLM calls per generation. Batch API calls and use smaller, faster models for evaluation if budget is constrained.

4. 4

   Select parents and apply crossover

   Select the top N candidates by fitness. For each new offspring, sample two parents and prompt an LLM to produce a crossover: "Combine the strongest elements of these two prompts to create a new variant that may perform better on [task]." The LLM performs semantically intelligent recombination rather than random text splicing.

5. 5

   Apply mutation to selected candidates

   Apply mutation to a subset of the population by prompting an LLM to make 1–3 targeted modifications to a candidate prompt (rephrase an instruction, add a constraint, change the output format, etc.). Mutation maintains diversity and prevents premature convergence to a local optimum.

6. 6

   Repeat for N generations and select the best

   Run 5–20 generations depending on budget. Track the best fitness score achieved across all generations. When improvement plateaus for 3+ generations, stop. Validate the best evolved prompt on a held-out test set (not the dev set used for fitness evaluation) to confirm the improvement generalizes.

Prompt Examples

## EvoPrompt — Manual Simulation (Generation 1)

### Task
Sentiment classification: classify customer reviews as Positive, Negative, or Neutral.

### Initial Population (3 candidates for illustration)

**Candidate A:**
Classify the sentiment of the following review as Positive, Negative, or Neutral.
Review: {review}
Sentiment:

**Candidate B:**
You are a sentiment analysis expert. Read the review carefully and determine
whether the overall customer sentiment is Positive, Negative, or Neutral.
Consider the overall impression, not just individual words.
Review: {review}
Answer with one word:

**Candidate C:**
Task: Sentiment Analysis
Instructions: Analyze the emotional tone of the customer review below.
Output exactly one of: Positive | Negative | Neutral
Do not explain your answer.
Review: {review}
Sentiment:

### Fitness Evaluation (accuracy on 50-example dev set)
- Candidate A: 74%
- Candidate B: 81%
- Candidate C: 78%

### Crossover Operation (LLM combines B and C)
Prompt the LLM: "Below are two high-performing prompt variants for a sentiment
classification task. Create a new improved variant that combines the strongest
elements of both. Keep the best structural and instructional elements from each."

[Insert Candidate B and Candidate C]

### Mutated Offspring (Generation 2 candidate)
You are a sentiment analysis expert. Analyze the emotional tone of the customer
review below and determine whether the overall impression is Positive, Negative,
or Neutral. Consider the full context, not isolated words.
Output exactly one of: Positive | Negative | Neutral
Review: {review}
Sentiment:

## EvoPrompt — Mutation Prompt Template

You are a prompt optimization assistant implementing the mutation operation
of an evolutionary algorithm for prompt optimization.

Below is a prompt that achieves {fitness_score}% accuracy on a {task_name} task.
Your goal is to produce a MUTATED variant that may perform better.

Mutation guidelines:
- Make 1–3 targeted changes, not a complete rewrite
- Valid mutations include: rephrasing an instruction, adding a constraint,
  changing the output format specification, adding or removing an example,
  adjusting the persona or tone, reordering sections
- Preserve elements that are likely contributing to performance
- Introduce one novel element not present in the original

Original prompt:
---
{original_prompt}
---

Output ONLY the mutated prompt text. No explanation, no commentary.
The mutated prompt should be immediately usable as a replacement.

Pros and Cons

Frequently Asked Questions