Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution

Chrisantha Fernando; Dylan Banarse; Henryk Michalewski; Simon Osindero; Tim Rocktäschel

Source-linked AI summary

Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution

Chrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, Tim Rocktäschel

arXiv:2309.16797v1Sep 28, 2023cs.CLcs.AIcs.LGcs.NE

TL;DR

Hand-crafted prompt strategies can be sub-optimal, motivating automated adaptation for different domains. Promptbreeder evolves task-prompts and the mutation-prompts that govern them, outperforming Plan-and-Solve prompting on reasoning benchmarks and producing complex prompts for hate speech classification.

Problem
Existing prompt strategies are manually engineered, although prompt phrasing can substantially affect foundation-model capabilities.
Method
Promptbreeder uses an LLM-driven evolutionary process to jointly evolve domain-specific task-prompts and mutation-prompts across generations.
Results
Promptbreeder outperforms Plan-and-Solve prompting across the reported reasoning benchmarks, with further few-shot gains, and evolves complex prompts for ETHOS hate speech classification.
Takeaways & Limitations
Promptbreeder demonstrates that self-referential evolution can automatically improve both task-prompts and the process used to modify them.
Takeaways & Limitations
Promptbreeder adapts prompt content within a fixed prompting topology rather than evolving the prompting algorithm itself.

Abstract

from arXiv · show

Popular prompt strategies like Chain-of-Thought Prompting can dramatically improve the reasoning abilities of Large Language Models (LLMs) in various domains. However, such hand-crafted prompt-strategies are often sub-optimal. In this paper, we present Promptbreeder, a general-purpose self-referential self-improvement mechanism that evolves and adapts prompts for a given domain. Driven by an LLM, Promptbreeder mutates a population of task-prompts, and subsequently evaluates them for fitness on a training set. Crucially, the mutation of these task-prompts is governed by mutation-prompts that the LLM generates and improves throughout evolution in a self-referential way. That is, Promptbreeder is not just improving task-prompts, but it is also improving the mutationprompts that improve these task-prompts. Promptbreeder outperforms state-of-the-art prompt strategies such as Chain-of-Thought and Plan-and-Solve Prompting on commonly used arithmetic and commonsense reasoning benchmarks. Furthermore, Promptbreeder is able to evolve intricate task-prompts for the challenging problem of hate speech classification.

1 INTRODUCTION

Promptbreeder automates prompt engineering through a diversity-maintaining evolutionary algorithm that self-referentially improves both task-prompts and the mutation-prompts used to evolve them. The paper reports improvements over established prompting strategies across arithmetic, commonsense reasoning, and ethics benchmarks without parameter updates.

Motivation: The method addresses diminishing returns in iterative prompt selection with a diversity-maintaining evolutionary algorithm for automated prompt engineering.This motivation follows observations that prompt phrasing strongly affects utility and that APE’s quality stabilized after three selection rounds.
Motivation: Promptbreeder frames prompts as an LLM’s program, allowing the model to modify both its prompting strategy and the process that modifies it.This extends the idea of self-referential self-improvement while avoiding the additional parameters required to modify all model weights.
Method and contributions: Promptbreeder evolves domain-adaptive task-prompts while simultaneously improving the mutation-prompts that generate them, enabling self-referential prompt improvement.It starts from seed mutation-prompts, thinking-styles, and a domain-specific problem description, then selects evolutionary units based on task-prompt fitness.
Results and contributions: The paper reports improvements over Chain-of-Thought and Plan-and-Solve prompting across commonly used arithmetic, commonsense reasoning, and ethics benchmarks.The approach requires no parameter updates for self-referential self-improvement.
Method and contributions: Promptbreeder uses binary tournament selection and five mutation-operator classes to evolve task-prompts and mutation-prompts across generations.Task-prompt fitness is measured on random training batches, producing increasingly domain-adaptive task-prompts and more useful mutation-prompts.

2 RELATED WORK

Promptbreeder builds on prior prompt-engineering and self-improvement research while addressing their limitations through automated, domain-adaptive evolution of natural-language prompts. It differs from existing methods by evolving both task-prompts and mutation-prompts without direct parameter updates or human intervention.

Prompting: Promptbreeder extends prompting strategies such as CoT, Zero-shot CoT, Self-Consistency, ToT, GoT, Plan-and-Solve, Least-to-Most, and Self-Refine by automatically adapting prompts to a target domain.These approaches improve reasoning through intermediate steps, sampling, search structures, planning, decomposition, or iterative feedback, but are generally hand designed and domain agnostic.
Prompting: Compared with APE, Promptbreeder evolves mutation-prompts online and uses population-aware operators, elite history, compositional initialization, and diversity maintenance to avoid diminishing returns and diversity loss.APE uses separate generator- and mutation-prompts to produce and mutate prompt candidates, whereas Promptbreeder improves the mutation process itself.
Prompting: Promptbreeder differs from OPRO and EvoPrompt by evolving multiple LLM-generated mutation-prompts alongside task-prompts, evaluating random training subsets, and self-referentially improving mutations.OPRO uses a single complex mutation prompt and a small fixed training set, while EvoPrompt uses fixed mutation and crossover prompts and a hand-designed task-prompt population.
Self-Referential Self-Improvement: Unlike parameter-based soft prompting and self-referential weight-update methods, Promptbreeder uses natural language as the substrate for self-referential improvement, avoiding costly model-parameter updates.Parameter updates may not scale to larger models or API-accessed LLMs, motivating a prompt-level alternative.
Open-Endedness and LLMs: Promptbreeder connects open-ended evolution with LLM capabilities for mutation and novelty, exploring prompts autonomously rather than evolving images or relying on humans in the loop.It exemplifies a transition from “learning from data” to “learning what data to learn from” (Jiang et al., 2022).

3 PROMPTBREEDER

Promptbreeder is a general-purpose evolutionary system that discovers domain-specific task-prompts while recursively improving the mutation-prompts that generate them. It combines population-based prompt evolution with diverse mutation, hypermutation, Lamarckian, crossover, and few-shot-context operators.

3 PROMPTBREEDER: Promptbreeder evolves task-prompts for a domain and evaluates each on batches of 100 training-set question–answer pairs.A task-prompt conditions an LLM before an input question to improve its response.
3 PROMPTBREEDER: Each task-prompt is mutated by an LLM conditioned on a mutation-prompt, while a hyper-mutation prompt evolves those mutation-prompts recursively.The system uses P′ = LLM(M + P) and M′ = LLM(H + M), with binary tournament replacement guiding population evolution.
3.1 PROMPTBREEDER INITIALIZATION: Initialization combines domain descriptions, randomly sampled mutation-prompts, and thinking styles to generate diverse task-prompts and associated mutation-prompts.For GSM8K, each evolutionary unit contains two task-prompts because Plan-and-Solve uses two, with sequential application producing the final answer.
3.2 MUTATION OPERATORS: Nine mutation operators across five classes explore alternative cognitive framings and natural-language mental models, with one operator sampled uniformly per replication.The design aims to recreate the role of diverse representational redescription in problem solving through self-referential prompt evolution.
3.2.1 DIRECT MUTATION: Direct mutation generates task-prompts either from existing parents or independently from the domain description, allowing standard variation and re-sampling when evolution diverges.First-order mutation is the standard asexual operator, whereas zero-order generation extracts a new hint without relying on prior task-prompts.
3.2.2 ESTIMATION OF DISTRIBUTION MUTATION: Distribution-based operators present filtered or lineage-ordered populations to the LLM, encouraging diverse extrapolations while withholding fitness values because preliminary tests produced copies.EDA filtering removes prompts exceeding 0.95 BERT embedding cosine similarity, and rank-based mutation uses ascending fitness order with a deliberately contradictory descending-order instruction.
3.2.3 HYPERMUTATION: MUTATION OF MUTATION-PROMPTS: Hypermutation evolves the search process itself by generating new mutation-prompts, including zero-order creation from thinking styles and first-order improvement of existing mutation-prompts.These operators target evolvability rather than directly optimizing task rewards, then apply the resulting mutation-prompts to downstream task-prompts.
3.2.4 LAMARCKIAN MUTATION; 3.2.5 PROMPT CROSSOVER AND CONTEXT SHUFFLING: Lamarckian operators reverse-engineer task-prompts from successful workings out, while crossover and context shuffling recombine task-prompts and evolve correct few-shot reasoning examples.Crossover replaces a task-prompt with 10% probability, and context shuffling replaces correct workings out after evaluation or resamples the full context with 10% probability.

4 EXPERIMENTS

The experiments evaluate Promptbreeder across arithmetic and commonsense reasoning, instruction induction, and hate speech classification using established benchmark datasets.

Experimental Setup: Promptbreeder was evaluated on arithmetic, commonsense reasoning, instruction induction, and hate speech classification benchmarks.The evaluation used a population of 50 units evolved for typically 20–30 generations, with random pairwise competitions each generation.

5 RESULTS AND DISCUSSION

Promptbreeder outperforms established prompting strategies across reasoning benchmarks, improves substantially on ETHOS hate-speech classification, and depends on its self-referential mutation mechanisms. The results also show that mutation operators continue improving task prompts throughout evolution.

Reasoning benchmarks: Promptbreeder achieves higher zero-shot accuracy than PS+ across the reported reasoning datasets, with further gains from few-shot discovered solutions.These comparisons use PS+PaLM 2-L on all datasets except ADDSUB, and the benchmark results are reported in Table 1.
ETHOS hate-speech classification: 89% on ETHOS versus 80% for the hand-designed baseline demonstrates Promptbreeder’s ability to evolve intricate, domain-adapted prompt strategies.The evolved strategy uses two sequentially applied relatively long prompts.
Mutation analysis: All mutation operators contribute to Promptbreeder, including hypermutation operators that enable self-referential self-improvement.Mutation-prompts are evaluated by how often they produce better task-prompts, while operator results are summarized in Appendix J.3–J.4.
Ablation study: Removing any self-referential operator is harmful in nearly all tested settings, although random mutation-prompt initialization specifically hurts GSM8K.The largest benefit comes from initially re-describing task-prompts.

6 CONCLUSION AND FUTURE WORK · A GLOSSARY

Promptbreeder is presented as a self-referential system that evolves both task-prompts and the mutation-prompts governing their improvement, while remaining limited by fixed prompting topology. The paper outlines future directions toward more open-ended, multimodal, and language-scalable self-improvement, and defines its core evolutionary terminology.

6 CONCLUSION AND FUTURE WORK: Promptbreeder evolves domain-specific task-prompts and the mutation-prompts that govern how those task-prompts are improved.Its self-referential mechanism improves both prompts and the process used to improve prompts.
6 CONCLUSION AND FUTURE WORK: Future work could evolve prompt diversity, evaluate whole conditional thought processes, and use self-play to evolve pre-prompts for competing LLM-based policies.The proposed examples include N-prompt strategies and competitive Socratic dialogue.
6 CONCLUSION AND FUTURE WORK: PB remains limited because it adapts prompt content while keeping the prompting topology fixed, unlike potentially reconfigurable, hierarchical, multimodal thought processes.The paper identifies language, intonation, and imagery as aspects of human thought not captured by the current system.
6 CONCLUSION AND FUTURE WORK: PB suggests that increasingly open-ended self-referential systems could use language as the substrate for improvement without parameter updates and scale with more capable LLMs.This is presented as a prospective direction rather than a demonstrated result.
A GLOSSARY: An estimation of distribution algorithm iteratively refines a probabilistic model of promising solutions, while fitness-proportionate selection chooses individuals according to population fitness.The glossary also calls fitness-proportionate selection roulette-wheel selection.
A GLOSSARY: A mutation prompt is concatenated with a task-prompt to generate an improved task-prompt, beginning from the user’s initial problem description.The problem description serves as Promptbreeder’s starting task-prompt.
A GLOSSARY: A prompt strategy comprises task-prompts and inference-time application rules, while phenotype, workings out, context, and reasoning path denote the LLM output for a prompted problem.In the minimal case, a strategy has one task-prompt; typical strategies use two sequential task-prompts.
A GLOSSARY: A population is the set of evolutionary units, each containing task-prompts, a mutation-prompt, and, in few-shot settings, two or three contexts.The glossary gives 50 as an example population size and describes a typical task-prompt set as two prompts.

B A TYPICAL EVOLUTIONARY RUN … H LAMARCKIAN MUTATION EXAMPLE

The merged sections show how Promptbreeder evolves task and mutation prompts, using diverse thinking styles and self-referential pathways to improve problem-solving prompts. They also illustrate the approach through a word-in-context run, generated prompt examples, task descriptions, and Lamarckian mutation.

B A TYPICAL EVOLUTIONARY RUN: Over 2000 evaluations, fitness increased considerably as two sequential prompts evolved for APE’s word-in-context task, with the best prompt pair reported.The prompts ask whether a homograph has the same sense or meaning in two sentences, and Figure 3 plots individual evaluations and population-mean fitness.
C MUTATION PROMPTS: Mutation prompts span rephrasing, creativity, unconventional perspectives, added specificity, problem decomposition, error correction, and other ways to transform instructions.The listed mutators include unexpected twists, contradictions, analogies, multiple perspectives, contextualization, visualization, reverse thinking, and correction prompts.
D THINKING STYLES: Thinking-style prompts frame problem solving through critical, creative, collaborative, systems, risk, reflective, analytical, design, and step-by-step approaches.They also ask about assumptions, obstacles, stakeholders, resources, evaluation metrics, alternative solutions, and modifying or replacing the current best solution.
E INITIALLY EVOLVED PROMPTS: Initially evolved prompts are generated by concatenating a thinking style, a mutation prompt, and the problem description.Table 4 provides examples of these initial prompts for GSM8k.
F PROMPTBREEDER AS SELF-REFERENTIAL SELF-IMPROVEMENT SYSTEM: Promptbreeder is self-referential because prompts, mutation prompts, offspring, inference workings, and Lamarckian mutations causally influence subsequent prompt generation through the LLM.The process can diverge or become trapped in attractors, and it evolves mutation methods but retains an externally specified fitness function rather than evolving evaluators.
G PROBLEM DESCRIPTIONS: Problem descriptions specify task outputs across arithmetic, multiple-choice mathematics, hate-speech classification, and commonsense reasoning benchmarks.Examples require Arabic-numeral answers, selecting options A–E, determining whether text contains hate speech, or answering commonsense questions with yes or no.
H LAMARCKIAN MUTATION EXAMPLE: In Lamarckian mutation, the LLM receives an instruction alongside example workings-out and generates a new prompt continuation based on those solutions.The example supplies stepwise advice and solved multiple-choice arithmetic problems, after which the model continues with a reusable “Solve like a pro!”-style prompt.

I DATASETS … I.5 INSTRUCTION INDUCTION

The paper evaluates Prompt Evolution across arithmetic, commonsense, hate-speech, and language-understanding tasks, with control prompts and evolved arithmetic prompts documented separately. These datasets span diverse reasoning, classification, and instruction-following challenges.

I.1 CONTROL TASK-PROMPTS: The arithmetic controls include Chain-of-Thought, Plan and Solve, Plan and Solve PS+, Zero-shot APE, and OPRO task-prompts.Table 5 lists the prompts, including a zero-shot APE prompt generated to improve over CoT on MultiArith and GSM8K.
I.2 ARITHMETIC REASONING: Prompt Evolution is evaluated on six arithmetic reasoning datasets spanning grade-school word problems, algebra, and multi-operation quantitative reasoning.The datasets are GSM8K, SVAMP, MultiArith, AddSub, AQuA-RAT, and SingleEq.
I.3 COMMONSENSE REASONING: Commonsense reasoning is evaluated on CommonsenseQA and StrategyQA, which require commonsense knowledge or multistep reasoning.CommonsenseQA uses multiple-choice questions, whereas StrategyQA contains yes/no questions requiring multiple reasoning steps.
I.4 HATE SPEECH CLASSIFICATION: The hate-speech experiment optimizes a long prompt for the ETHOS dataset, following a task previously attempted by Pryzant et al. (2023).That work used workingout-conditioned error detection and error-fixing prompts, a self-referential process similar to Prompt Evolution’s Lamarckian operator.
I.5 INSTRUCTION INDUCTION: Instruction Induction contains 24 language-understanding tasks ranging from spelling and morphosyntax to similarity, causality, style transfer, and sentiment analysis.The tasks vary in difficulty and include examples such as pluralization and formality transformation.
I.5 INSTRUCTION INDUCTION: The study also documents two-stage task-prompts evolved for different arithmetic tasks.These prompts are presented in Table 6.

J EXAMPLE RESULTS … J.4 MUTATION OPERATOR EFFECTIVENESS

The example results show Promptbreeder evolving hate-speech task prompts and self-referential mutation prompts, with experiments using defined population, fitness, sampling, stopping, and diversity procedures. The appendices also characterize mutation-prompt success and rank mutation operators by offspring-improvement frequency on GSM8K.

J.1 ETHOS EVOLVED PROMPT: Promptbreeder evolved two hate-speech classification prompts, including criteria for derogatory, generalized, threatening, discriminatory, or context-dependent language.The prompts define target groups, harmful speech, intent, audience, and context, while noting humorous or satirical language may not constitute hate speech.
J.2 PROMPT EVOLUTION MATHS RESULTS: The experiments used populations of 50, with fitness measured as accuracy on randomly selected batches of 100 training examples.Datasets without predefined splits were divided equally into training and test sets before experimentation.
J.2 PROMPT EVOLUTION MATHS RESULTS: Promptbreeder sampled the LLM as Redescriber, Inducer, and Evaluator, with maximum outputs of 50, 30, and 5 tokens, respectively.Inducer and Evaluator temperatures were fixed at 0.0, while Redescriber temperature evolved from an initialization range of 1.0 to 2.0.
J.2 PROMPT EVOLUTION MATHS RESULTS: Runs continued until training fitness plateaued, after which the fittest individual was evaluated on the test set.Experiments generally used 1–2k fitness evaluations, corresponding to roughly 20–40 generations for 25 pair evaluations per generation.
J.2 PROMPT EVOLUTION MATHS RESULTS: Three diversity mechanisms addressed local optima: random character prefixes, BERT-similarity fitness sharing, and mutation of the Redescriber’s sampling temperature.The Redescriber temperature was initialized uniformly between 1.0 and 2.0 and mutated by a uniform random value from -0.2 to 0.2 at each replication.
J.2 PROMPT EVOLUTION MATHS RESULTS: Comparisons with PoT, PS, and Auto-CoT controls were not provided because PS and PS+ were the best Plan-and-Solve prompts.This limits the reported control comparison for the mathematics experiments.
J.3 EVOLVED MUTATION PROMPTS: Table 7 lists mutation prompts evolved self-referentially during GSM8K training and defines each score as the probability of producing an improved prompt.The table identifies the most successful evolved mutation prompts rather than reporting task-prompt results.
J.4 MUTATION OPERATOR EFFECTIVENESS: Table 8 ranks GSM8K mutation operators by the proportion of offspring whose fitness exceeded their parent’s fitness.The supplied passage identifies the table’s scope and ranking criterion but provides no operator-specific proportions.

J.5 ADDSUB

After 1600 mutations, few-shot evolution produced task-prompts that often drifted into nonsense, while evolved contexts dominated fitness. The evolved contexts could still support arithmetic answers despite containing unrelated or corrupted text.

ADDSUB: After 1600 mutations, evolved contexts were more critical to fitness than task-prompts, which often drifted into nonsense during few-shot evolution.The procedure applies Prompt 0 to produce a working-out and concatenates it with Prompt 1 to produce the answer, as in Plan-And-Solve.
ADDSUB: Despite corrupted context, one evolved response correctly computed 6689 - 660 = 6029 gallons of blue paint.The context included unrelated reading-comprehension material, while the response retained the relevant arithmetic operation and answer.
ADDSUB: Other evolved responses correctly answered seashell problems with 21 total seashells and 3 unbroken seashells amid nonsensical game instructions.The calculations used 6 + 8 + 7 = 21 and 7 - 4 = 3, although the surrounding generated text was corrupted.

J.6 AQUA … J.10 SVAMP

The appendix shows evolved prompts and examples across AQUA, MULTIARITH, GSM8K, SINGLEEQ, and SVAMP, illustrating domain-specific instructions and generated solutions. The prompts range from explicit answer-format and solution instructions to highly specialized or noisy mutations.

J.6 AQUA: On AQUA, evolved prompts instruct the model to understand multiple-choice math word problems before solving and selecting an option.The displayed prompts explicitly require choosing among (A), (B), (C), (D), or (E).
J.6 AQUA: AQUA examples apply these instructions to ratio, travel-time, and simple-interest problems, producing answers such as (A), (E), and (B).The examples calculate 12,000 and 20,000 for the ratio problem, identify (E) for the lizard problem, and obtain Rs 8,925 for the interest problem.
J.7 MULTIARITH: MULTIARITH prompts combine Arabic-numeral answer formatting with recalling and reusing solutions to similar problems.The evolved instructions ask the model to solve the word problem as an Arabic numeral and ask whether it has solved a similar problem before.
J.7 MULTIARITH: MULTIARITH examples demonstrate this pattern on addition, subtraction, and multiplication word problems, yielding 22, 1, and 32.The examples respectively compute Frank’s remaining money, Emily’s final score, and Maria’s furniture-assembly time.
J.8 GSM8K: GSM8K evolves a concise instruction to solve math word problems without a calculator and provide the answer as an Arabic numeral.The accompanying examples solve sibling-count, bracelet-sales, and cookie-cost problems with answers 135, 132, and 3.
J.9 SINGLEEQ / J.10 SVAMP: The passages show prompts containing quotation text, inverse-variation instructions, and a date range, alongside examples stating direct numerical answers of 426, 21, and 12.
J.10 SVAMP: SVAMP examples solve pages-per-day, combined-email, and remaining-cookie problems through straightforward arithmetic, returning 102, 11, and 1.The generated solutions divide 612 pages by 6 days, add 3 and 8 emails, and subtract 36 cookies from 37.

K APE INSTRUCTION INDUCTION TASKS · Q. 123456789 A. Write out the first letter of each input.

This section presents instruction-induced prompts and outputs across diverse tasks, including reasoning, rewriting, arithmetic, translation, sentence similarity, and word transformations. The examples illustrate task-specific procedures and direct answers, alongside Promptbreeder’s performance relative to APE across instruction-induction tasks.

K APE INSTRUCTION INDUCTION TASKS: Promptbreeder was initialized with a randomly selected induction input example because these datasets lack explicit problem descriptions.APE likewise uses induction input examples, consisting of a fixed prompt and several training examples from each task.
K APE INSTRUCTION INDUCTION TASKS: Up to 21% improvement over APE was achieved on 21 of 24 instruction-induction tasks, despite Promptbreeder using an instruction-untuned LLM.Table 9 compares PaLM2-L Prompt Evolution with three APE controls, including published text-davinci-002 results and an APE-initialized Promptbreeder variant.
K.1.1 FIRST LETTER / K.1.2 SECOND LETTER: For first-letter and second-letter tasks, evolved prompts instructed character extraction, while contexts demonstrated outputs such as the first letter of “drummer” and the second letters of “month,” “asylum,” and “heel.”The first-letter prompt explicitly says to write the first letter of each input; second-letter contexts reason through input length and select the second character.
Q. 123456789 A. Write out the first letter of each input.: The instruction-mutant example applies “Write out the first letter of each input” to produce answers including “r” and a result paired with 12.A related context gives “rest” while the examples retain the instruction-mutant framing.
K.1.3 LIST LETTERS / K.1.4 STARTING WITH: For list-letter and starting-with tasks, evolved contexts specified inserting spaces after consonants or returning the first word containing a marked letter.Examples include “g a l e,” “a c c o m p a n i m e n t,” and returning “impressed” for a sentence marked with [i].
K.1.5 PLURALIZATION: Pluralization prompts evolved an explicit y-to-ies rule and otherwise added s, producing “touches,” “forages,” and “miles.”The same evolved instruction is shown with all three inputs and outputs.
K.1.6 PASSIVIZATION / K.1.7 NEGATION: Passivization and negation contexts encoded grammatical transformations, producing passive sentences such as “The presidents were stopped by the authors” and negations such as “Some people are not wise.”The passivization prompt swaps noun phrases and conjugates the verb, while negation examples insert “not” or use “did not.”
K.1.8 ANTONYMS / K.1.9 SYNONYMS / K.1.10 MEMBERSHIP: Antonym, synonym, and membership examples evolved outputs including “partisan,” “dignified,” “unflattering,” synonym lists for “bus” and “frightened,” and alphabetically ordered animal subsets.The membership examples exclude nonanimals and return ordered lists such as “goat, penguin, shark, lion” and “rabbit, snail, jellyfish.”
K.1.5 PLURALIZATION / K.1.6 PASSIVIZATION / K.1.7 NEGATION / K.1.8 ANTONYMS / K.1.9 SYNONYMS / K.1.10 MEMBERSHIP / K.1.11 RHYMES: Additional evolved prompts covered pluralization, passivization, negation, antonyms, synonyms, membership, and rhyming transformations, including removing a final e or replacing the last two letters with “mote.”The rhyming subsection presents rules for modifying final letters, while the other subsections provide task-specific transformation prompts and examples.
K.1.11 RHYMES: The RHYMES task applies letter-removal and suffix rules to inputs such as “pea,” “night,” and “add,” producing the stated answers.For “pea,” the passage states that the correct answer is “pea”; analogous examples state “night” and “add.”
K.1.12 LARGER ANIMAL: The cited examples compare animal weights and identify the heavier animal: sperm whale over sheltie (450000 versus 25 pounds), and polar bear over hummingbird (1500 versus 0.002 pounds).
K.1.13 CAUSE SELECTION: The CAUSE SELECTION task asks which sentence caused an effect or happened first, yielding bottle left open before soda went flat, wind before capsizing, and power loss before spoiled meat.These answers are explicitly given for the three example pairs.
K.1.14 FORMALITY: The FORMALITY task rewrites informal statements into formal language, producing “The entire story was fabricated” and “It is my opinion that this is interesting.”The examples repeatedly specify formal rewrites for both source sentences.
K.1.15 SUM / K.1.16 DIFFERENCE: The SUM and DIFFERENCE tasks demonstrate arithmetic instructions, with examples returning 140, 52, and 41 for sums and 44, 48, and 49 for differences.The difference examples explicitly compute 73−29=44, 73−25=48, and 53−4=49.
K.1.18 TRANSLATION ENGLISH-GERMAN: The TRANSLATION ENGLISH-GERMAN task translates words such as “relation” and “camp” as “Beziehung” and “Lager,” while “lecture” is given as “Vortrag.”The examples state these German answers directly.
K.1.19 TRANSLATION ENGLISH-SPANISH / K.1.20 TRANSLATION ENGLISH-FRENCH: The English-Spanish and English-French tasks provide direct word translations, including “bear”→“oso,” “relativity”→“relatividad,” “destiny”→“destin,” and “ideology”→“ideologie.”The Spanish examples also translate concert-related words, including “concert” and “concerto” as “concierto.”
K.1.20 TRANSLATION ENGLISH-FRENCH: The translation example asks for five English-to-French translations and concludes with the answer “representation.”The displayed prompt includes repeated representation entries and a concluding translation instruction.
K.1.22 SENTENCE SIMILARITY: For sentence similarity, the algorithm must determine similarity between two sentences on a scale of 0 to 5; 0 means nothing in common, while 5 means semantically similar.
K.1.22 SENTENCE SIMILARITY: The sentence-similarity examples assign score 1 when sentences concern the same topic but express different ideas.This outcome appears for examples about an Iranian nuclear plant and a girl and band playing.
K.1.23 WORD IN CONTEXT: The word-in-context prompts ask whether a target word has the same meaning in two sentences and require a yes-or-no response.The prompt wording appears in both Prompt 0 and Prompt 1.
K.1.23 WORD IN CONTEXT: The word-in-context examples label “voice” and “communicate” as same-meaning uses, producing yes answers.Both examples conclude with the answer yes.
K.1.23 WORD IN CONTEXT: The word “take” is labeled as having different meanings across “take me to the main entrance” and “Take a scene,” producing no.The example explicitly concludes with the answer no.

K.1.21 SENTIMENT ANALYSIS

The sentiment-analysis section presents prompts for classifying reviews by positive or negative language and for counting positive examples. Example applications show both prompt formats producing positive sentiment answers for movie reviews.

The primary task prompt asks whether each review uses generally positive or negative language.
A second prompt instead asks the model to write the number of positive examples.
Example outputs classify reviews such as “Gangs” and “Read My Lips” as positive under the sentiment prompts.The supplied examples include the review “I simply can’t recommend it enough” and indicate a positive answer.

L ABLATIONS

Ablations show that Promptbreeder’s self-referential mutation operators generally improve fitness, with random initial prompts having the largest positive impact across datasets. On ETHOS, removing Lamarckian context mutation causes the greatest performance drop, demonstrating interactions among the operators.

L ABLATIONS: The ablations removed random initial prompts, random initial mutation prompts, Lamarckian context-to-task-prompt mutation, and meta-mutation of mutation-prompts.For Lamarckian mutation and meta-mutation, the default zero-/first-order mutation operation was used with a 50:50 chance.
L ABLATIONS: Across datasets, most mutation operators positively affected fitness, with random initial prompts producing the largest positive impact.Each ablation used a population of 10 for 200 evaluations and was compared with the complete algorithm at the same population size.
L ABLATIONS: 81.6% was achieved on ETHOS, falling to 64.6% without Lamarckian context mutation and to 68.7% when random initialization and hyper-mutation were also removed.Online mutation restored performance to 70.4%, while adding random mutation prompts increased it to 73.7%.