Published as a conference paper at ICLR 2023

L ARGE L ANGUAGE M ODELS ARE H UMAN -L EVEL

P ROMPT E NGINEERS
Yongchao Zhou1,2,∗ , Andrei Ioan Muresanu2,3,∗ , Ziwen Han1,2,∗ , Keiran Paster1,2 ,
Silviu Pitis1,2 , Harris Chan1,2 , Jimmy Ba1,2
1
University of Toronto 2 Vector Institute 3 University of Waterloo ∗ Equal contribution
{yczhou,hanziwen,keirp,spitis,hchan,jba}@cs.toronto.edu
{andrei.muresanu}@uwaterloo.ca
arXiv:2211.01910v2 [cs.LG] 10 Mar 2023

A BSTRACT
By conditioning on natural language instructions, large language models (LLMs) have
displayed impressive capabilities as general-purpose computers. However, task performance
depends significantly on the quality of the prompt used to steer the model, and most effective
prompts have been handcrafted by humans. Inspired by classical program synthesis and
the human approach to prompt engineering, we propose Automatic Prompt Engineer1
(APE) for automatic instruction generation and selection. In our method, we treat the
instruction as the “program,” optimized by searching over a pool of instruction candidates
proposed by an LLM in order to maximize a chosen score function. To evaluate the
quality of the selected instruction, we evaluate the zero-shot performance of another LLM
following the selected instruction. Extensive experiments show that our automatically
generated instructions outperform the prior LLM baseline by a large margin and achieve
better or comparable performance to the instructions generated by human annotators on
24/24 Instruction Induction tasks and 17/21 curated BIG-Bench tasks. We conduct extensive
qualitative and quantitative analyses to explore the performance of APE. We show that
APE-engineered prompts are able to improve few-shot learning performance (by simply
prepending them to standard in-context learning prompts), find better zero-shot chain-of-
thought prompts, as well as steer models toward truthfulness and/or informativeness. 2

1 I NTRODUCTION
The combination of scale and attention-based architectures has resulted in language models possessing
an unprecedented level of generality (Kaplan et al., 2020; Vaswani et al., 2017). These so-called
“large language models” (LLMs) have shown remarkable, often superhuman, capabilities across a
diverse range of tasks, including both zero-shot and few-shot setups (Brown et al., 2020; Srivastava
et al., 2022). With generality, however, there comes a question of control: how can we make LLMs
do what we want them to do?
To answer this question and steer LLMs toward desired behaviors, recent work has considered
fine-tuning (Ouyang et al., 2022; Ziegler et al., 2019), in-context learning (Brown et al., 2020), and
several forms of prompt generation (Gao, 2021), including both differentiable tuning of soft prompts
(Qin & Eisner, 2021; Lester et al., 2021) and natural language prompt engineering (Reynolds &
McDonell, 2021). The latter is of particular interest, as it provides a natural interface for humans to
communicate with machines and may be of great relevance not only to LLMs but to other generalist
models such as prompted image synthesizers (Rombach et al., 2022; Ramesh et al., 2022), for which
public interest in prompt design and generation has also emerged (see Appendix A for examples).
Behind this interest is the fact that plain language prompts do not always produce the desired results,
even when those results are possible to produce with alternative instructions. Thus, human users must
experiment with a wide range of prompts to elicit desired behaviors, as they have little knowledge of
how compatible instructions are with a particular model. We can understand this by viewing LLMs
as black-box computers that execute programs specified by natural language instructions: while they
1
We define “prompt engineering” as optimizing the language in a prompt in order to elicit the best possible
performance. Notably, this does not include prompts that chain multiple LLM queries together or give the LLM
access to external tools.
2
Our code is available at https://github.com/keirp/automatic_prompt_engineer.

1
Published as a conference paper at ICLR 2023

Keep the high score candidates Discard the low score candidates Final selected prompt with highest score
0.81
0.8

Interquartile Mean Zero-Shot Performance

LLMs as Inference Models LLMs as Scoring Models
0.75 0.71 0.73
Professor Smith was given the Instruction: write the antonym of the
following instructions: <INSERT> word. <LIKELIHOOD>
Human Prompt Engineer 0.63
0.65
Here are the Professor’s responses: Input: direct Output: indirect 0.61
0.6 0.59
0.57
# Demostration Start
Input: prove Output: disprove
①
Proposal
② Scoring ③ Log
Probability
Input: on Output: off
... write the antonym of the word. -0.26
# Demostration End
give the antonym of the word provided. -0.28 0.4 0.40

④ ... ...
[Optional] High Score
reverse the input. -0.86
Candidates
LLMs as Resampling Models
to reverse the order of the letters -1.08 0.2
⑤
Generate a variation of the following
instruction while keeping the semantic
meaning. Similar write the opposite of the word given. -0.16
Candiates
Input: write the antonym of the word. 0.03 0.03
... ...
0.01 0.01 0.02 0.01 0.03
Output: <COMPLETE> list antonyms for the given word. -0.39 0
350M 1.3B 6.7B 175B 350M 1.3B 6.7B 175B 350M 1.3B 6.7B 175B 350M 1.3B 6.7B 175B
Greedy (GPT-3) Greedy (InstructGPT) APE (GPT-3) APE (InstructGPT)

(a) Automatic Prompt Engineer (APE) workflow (b) Interquartile mean across 24 tasks

Figure 1: (a) Our method, Automatic Prompt Engineer (APE), automatically generates instructions
for a task that is specified via output demonstrations: it generates several instruction candidates, either
via direct inference or a recursive process based on semantic similarity, executes them using the target
model, and selects the most appropriate instruction based on computed evaluation scores. (b) As
measured by the interquartile mean across the 24 NLP tasks introduced by Honovich et al. (2022),
APE is able to surpass human performance when using the InstructGPT model (Ouyang et al., 2022).

can execute a broad range of natural language programs, the way these programs are processed may
not be intuitive for humans, and the quality of instruction can only be measured when executing these
instructions on a downstream task (Sanh et al., 2022; Wei et al., 2021).
To reduce the human effort involved in creating and validating effective instructions, we propose a
novel algorithm using LLMs to generate and select instructions automatically. We call this problem
natural language program synthesis and propose to address it as a black-box optimization problem
using LLMs to generate and search over heuristically viable candidate solutions. In doing so, we
leverage the generalist capabilities of LLMs in three ways. First, we use an LLM as an inference
model (Ellis et al., 2021; Honovich et al., 2022) to generate instruction candidates based on a small set
of demonstrations in the form of input-output pairs. Next, we guide the search process by computing
a score for each instruction under the LLM we seek to control. Finally, we propose an iterative Monte
Carlo search method where LLMs improve the best candidates by proposing semantically similar
instruction variants. Intuitively, our algorithm asks LLMs to generate a set of instruction candidates
based on demonstrations and then asks them to assess which instructions are more promising. We
call our algorithm Automatic Prompt Engineer (APE). Our main contributions are:

• We frame instruction generation as natural language program synthesis, formulate it as a

black-box optimization problem guided by LLMs, and propose both a naive and an iterative
Monte Carlo search methods to approximate the solution.
• Our proposed method, APE, achieves human-level performance on zero-shot learning with
model-generated instructions on 24/24 Instruction Induction and 17/21 Big-Bench tasks.
• We provide extensive qualitative and quantitative analyses exploring various facets of
APE, and demonstrate applications of APE for improving few-shot learning, finding better
zero-shot chain of thought prompts, and steering LLMs toward desired behaviors such as
truthfulness and/or informativeness.

2 R ELATED W ORK
Large Language Models Scaling up transformer-based language models in terms of model size,
training data, and training compute has been shown to predictably improve performance on a wide
range of downstream NLP tasks (Vaswani et al., 2017; Devlin et al., 2018; Brown et al., 2020).
Many emergent abilities (Wei et al., 2022a) of LLMs have been discovered as a result of this scaling,
including few-shot in-context learning, zero-shot problem solving, chain of thought reasoning,
instruction following, and instruction induction (Cobbe et al., 2021; Wei et al., 2022b; Kojima et al.,

2
Published as a conference paper at ICLR 2023

2022; Sanh et al., 2022; Wei et al., 2021; Ouyang et al., 2022; Honovich et al., 2022). In this paper, we
view LLMs as black-box computers that execute programs specified by natural language instructions
and investigate how to control an LLM’s behavior using model-generated instructions.
Prompt Engineering Prompting offers a natural and intuitive interface for humans to interact
with and use generalist models such as LLMs. Due to its flexibility, prompting has been widely
used as a generic method for NLP tasks (Schick & Schütze, 2021; Brown et al., 2020; Sanh et al.,
2022). However, LLMs require careful prompt engineering, either manually (Reynolds & McDonell,
2021) or automatically (Gao et al., 2021; Shin et al., 2020), as models do not seem to understand the
prompts in the same way a human would (Webson & Pavlick, 2021; Lu et al., 2021). Though many
successful prompt tuning methods perform optimization over a continuous space using gradient-based
methods (Liu et al., 2021; Qin & Eisner, 2021; Lester et al., 2021), this becomes less practical with
scale, as computing gradients becomes increasingly expensive and access to models shifts to APIs
that may not provide gradient access. In our paper, we borrow components from discrete prompt
search methods, such as prompt generation (Gao et al., 2021; Ben-David et al., 2021), prompt scoring
(Davison et al., 2019) and prompt paraphrasing (Jiang et al., 2020; Yuan et al., 2021) to optimize
instructions by searching directly in the natural language hypothesis space. As compared to this past
work, which uses specialized models for each component and leans heavily on human templates, we
show that the entire search can be conducted by a single LLM.
Program Synthesis Program synthesis involves the automatic search over a “program space” to
find a program satisfying a particular specification (Gulwani et al., 2017). Modern program synthesis
admits a wide variety of specifications, including input-output examples (Ellis et al., 2021; Wong
et al., 2021) and natural language (Jain et al., 2022). The range of feasible program spaces to search
over has also grown, from historically restrictive domain-specific languages to general-purpose
programming languages (Austin et al., 2021). In contrast to prior approaches that require a suitable
structured hypothesis space and library of components (Liang et al., 2010; Ellis et al., 2018), we
leverage the structure provided by LLMs to search over the space of natural language programs.
Using inference models is a standard practice to speed up the search by restricting the search space to
a limited space of possible expressions (Menon et al., 2013; Lee et al., 2018; Devlin et al., 2017; Ellis
et al., 2021). Inspired by this, we use LLMs as approximate inference models to generate program
candidates based on a small set of demonstrations. Unlike classical program synthesis, our inference
models do not require any training and generalize well to various tasks.

3 NATURAL L ANGUAGE P ROGRAM S YNTHESIS USING LLM S

We consider a task specified by a dataset Dtrain = {(Q, A)} of input/output demonstrations sampled
from population X , and a prompted model M. The goal of natural language program synthesis
is to find a single instruction ρ such that, when M is prompted with the concatenation [ρ; Q] of
instruction and a given input, M produces the corresponding output A. More formally, we frame this
as an optimization problem, where we seek instruction ρ that maximizes the expectation of some
per-sample score f (ρ, Q, A) over possible (Q, A):
ρ? = arg max f (ρ) = arg max E(Q,A) [f (ρ, Q, A)] (1)
ρ ρ

Note that in general, Q may be the empty string, such that we are optimizing ρ as a prompt that
directly produces outputs {A}. While this task has been widely attempted by humans, we have little
knowledge of how compatible any particular instruction is with model M. Thus, we propose to treat
this human-intractable question as a black-box optimization process guided by LLMs. Our algorithm,
APE, uses LLMs in each of two key components, proposal and scoring. As shown in Figure 1 and
summarized in Algorithm 1, APE first proposes a few candidate prompts, and then filters/refines
the candidate set according to a chosen score function, ultimately choosing the instruction with the
highest score. We discuss options for proposal and scoring next.

3.1 I NITIAL P ROPOSAL D ISTRIBUTIONS

Due to the infinitely large search space, finding the right instruction can be extremely difficult, which
has rendered natural language program synthesis historically intractable. Recent progress in NLP
has shown language models are very good at generating diverse natural language text. Therefore, we

3
Published as a conference paper at ICLR 2023

Algorithm 1 Automatic Prompt Engineer (APE)

Require: Dtrain ← {(Q, A)}n : training examples, f : ρ × D 7→ R: score function
1: Use LLM to sample instruction proposals U ← {ρ1 , ..., ρm }. (See Section 3.1)
2: while not converged do
3: Choose a random training subset D etrain ⊂ Dtrain .
4: for all ρ in U do
5: Evaluate score on the subset se ← f (ρ, Detrain ) (See Section 3.2 )
6: end for
7: Filter the top k% of instructions with high scores Uk ⊂ U using {e s1 , ..., sem }
8: Update instructions U ← Uk or use LLM to resample U ← resample(Uk ) (See Section 3.3)
9: end while
Return instruction with the highest score ρ? ← arg maxρ∈Uk f (ρ, Dtrain )

consider leveraging a pretrained LLM to propose a good set U of candidate solutions that will guide
our search procedure. While random samples from LLMs are unlikely to produce the desired (Q, A)
pairs, we can instead ask the LLM to approximately infer the most likely instructions with a high score,
given the input/output demonstrations; i.e., to approximately sample from P (ρ | Dtrain , f (ρ) is high).
Forward Mode Generation We consider two approaches to gen- Forward Generation Template
erate high-quality candidates from P (ρ | Dtrain , f (ρ) is high). First,
I gave a friend an instruction and five
we adopt an approach based on “forward” mode generation by trans- inputs. The friend read the instruction
lating this distribution P (ρ | Dtrain , f (ρ) is high) into words. For and wrote an output for every one of
the inputs. Here are the input-output
example, in our instruction induction experiments (Subsection 4.1), pairs:
we follow Honovich et al. (2022) and prompt the LLM using Figure
Input: [ ] Output: [ ]
2 (Top). Input: [ ] Output: [ ]
...
Reverse Mode Generation Although the “forward” model works
The instruction was <COMPLETE>
out of the box for most of the pretrained LLMs, translating
P (ρ | Dtrain , f (ρ) is high) into words requires custom engineering
Reverse Generation Template
across different tasks. This is because while instructions are typi-
I instructed my friend to <INSERT>.
cally found in the beginning of passages, the “forward” model only
generates text from left to right, which requires the instruction to be The friend read the instruction and
wrote an output for every one of the
predicted at the end of the prompt. Therefore, we desire a more flex- inputs. Here are the input-output pairs:
ible approach such that the instruction can be anywhere in the text.
Input: [ ] Output: [ ]
To address this, we consider “reverse” mode generation, which uses Input: [ ] Output: [ ]
an LLM with infilling capabilities—e.g., T5 (Raffel et al., 2020), ...
GLM (Du et al., 2022), and InsertGPT (Bavarian et al., 2022)—to
infer the missing instructions. Our “reverse” model directly samples Template for TruthfulQA
from P (ρ | Dtrain , f (ρ) is high) by filling in the blank. We show an Professor Smith was given the
example of the such template in Figure 2 (Middle). following instructions: <INSERT>

Here are the Professor’s responses:

Customized Prompts Note that depending on the score function
Input: [ ] Output: [ ]
being used, there may exist more appropriate prompts than the sam- Input: [ ] Output: [ ]
ples above. For example, in our TruthfulQA experiments, we start ...
with the human-designed instructions from the original dataset (Lin
et al., 2022) and ask the the “reverse” model to propose initial in- Figure 2: Prompts for LLMs
struction samples that fit the missing context (Figure 2 (Bottom)).

3.2 S CORE F UNCTIONS

To cast our problem as black-box optimization, we choose a score function that accurately measures
the alignment between the dataset and the data the model generates. In our instruction induction
experiments, we consider two potential score functions, described below. In the TruthfulQA ex-
periments, we focused primarily on automated metrics proposed in Lin et al. (2022), similar to the
execution accuracy. In each case, we evaluate the quality of a generated instruction using Equation
(1), and take the expectation over a held-out test dataset Dtest .
Execution accuracy First, we consider evaluating the quality of an instruction ρ using the execution
accuracy metric proposed by Honovich et al. (2022), which we denote as fexec . In most cases,

4
Published as a conference paper at ICLR 2023

execution accuracy is simply defined as the 0-1 loss, f (ρ, Q, A) = 1 [M([ρ; Q]) = A]. On some
tasks, execution accuracy takes into account invariants; e.g., it may be an order invariant set matching
loss, as described in Appendix A of Honovich et al. (2022).
Log probability We further consider a softer probabilistic score function, which we hypothesize
might improve optimization by providing a more fine-grained signal when searching over low-quality
instruction candidates. In particular, we consider the log probability of the desired answer given the
instruction and question under the target model M, which on a per sample basis, is log P (A | [ρ; Q]).
Efficient score estimation Estimating the score by computing the score over the entire training
dataset for all instruction candidates can be expensive. To reduce the computation cost, we adopt
a filtering scheme where a promising candidate receives more computation resources while a low-
quality candidate receives less computation. It can be achieved by using a multi-stage computation
strategy on lines 2-9 Algorithm 1. We first evaluate all candidates with a small subset of the training
dataset. For the candidates with a score greater than a certain threshold, we sample and evaluate
a new non-overlapping subset from the training dataset to update the moving average of the score.
Then, we repeat this process until a small set of candidates is left, which are evaluated on the entire
training dataset. This adaptive filtering scheme significantly improves the computation efficiency
by keeping the exact computation costs for the high-quality samples and drastically reducing the
computation costs for low-quality candidates. We note that a similar score estimation scheme has
been used in previous works (Li et al., 2022; Maclaurin & Adams, 2015).

3.3 I TERATIVE P ROPOSAL D ISTRIBUTIONS

Despite our attempt to directly sample high-quality initial instruction candidates, it could be the case
that the method described in Subsection 3.1 fails to produce a good proposal set U, either because
it lacks of diversity or does not contain any candidates with a suitably high score. In case of such
challenges, we explore an iterative process for resampling U.
Iterative Monte Carlo Search Instead of only sampling from
Prompt for Resampling
the initial proposal, we consider exploring the search space locally
Generate a variation of the
around the current best candidates. This allows us to generate new following instruction while
instructions that are more likely to be successful. We call this variant keeping the semantic meaning.
iterative APE. At each stage, we evaluate a set of instructions and
Input: [INSTRUCTION]
filter out candidates with low scores. Then, an LLM is asked to
generate new instructions similar to those with high scores. We Output: <COMPLETE>
provide the prompt used for resampling in Figure 3. Figure 6 (Right)
shows that although this approach improves the overall quality of
Figure 3: Resampling
the proposal set U, the highest scoring instruction tends to remain
the same with more stages. We conclude iterative generation provides marginal improvement over the
relative simplicity and effectiveness of the generative process described in Subsection 3.1. Therefore,
we use APE without iterative search as default unless otherwise stated.

4 L ARGE L ANGUAGE M ODELS ARE H UMAN -L EVEL P ROMPT E NGINEERS

This section examines how APE can guide LLMs to desired behaviors. We investigate from four
perspectives: zero-shot performance, few-shot in-context learning performance, zero-shot chain-of-
thought reasoning, and truthfulness. Our experiments show that APE can find prompts that improve
task performance, performing equal to or even better than those authored by humans. APE also
often produces insightful tricks for how to best prompt language models that can be successfully
transferred to new tasks (see Section 4.3).

4.1 I NSTRUCTION I NDUCTION

We assess the effectiveness of zero-shot and few-shot in-context learning on 24 instruction induction
tasks proposed in Honovich et al. (2022). The tasks span many facets of language understanding, from
simple phrase structure to similarity and causality identification. We provide a detailed descriptions
of each task in Appendix B. For each task, we sample five input-output pairs from the training data
and select the best instruction using algorithm 1. Then, we evaluate the quality of the instruction

5
Published as a conference paper at ICLR 2023

Greedy Human APE

1

0
Execution Accuracy Antonyms Cause Selection Common Concept Diff First Letter Formality Large Animal List Letters
1

0
Membership Negation Number to Word Passivization Pluralization Rhymes Second Letter Sentence Similarity
1

0
Sentiment Starting With Sum Synonyms Translation en-de Translation en-es Translation en-fr Word in Context

Figure 4: Zero-shot test accuracy on 24 Instruction Induction tasks. APE achieves human-level or
better performance on all 24 out of 24 tasks.

by executing the instruction on InstructGPT 3 . We repeat our experiments five times with different
random seeds to report the mean and standard deviation. The exact templates for our experiments can
be found in Appendix (Table 5).
Zero-shot Learning We compare our method against two baselines: human prompt engineers
(Human)4 and the model-generated instruction algorithm proposed by Honovich et al. (2022). This
algorithm can be thought of as a greedy version of APE, without a search and selection process;
thus, we refer to it as “Greedy”. Figure 4 shows the zero-shot performance of InstructGPT using
human instructions and model generated instructions. Our algorithm outperforms “Greedy” on
every task and achieves equal or better than human performance on 24 of 24 tasks. Moreover, the
Interquartile Mean (IQM) (Agarwal et al., 2021) across all 24 tasks in Figure 1 suggests that APE with
InstructGPT outperforms human-engineered prompts, obtaining an IQM of 0.810 vs humans’ 0.749.
We summarize the instruction selected by APE for each task in Appendix (Table 12).
Few-shot In-context Learning We evaluated APE-generated instructions in few-shot in-context
learning, where we insert the instruction before the in-context demonstrations. Those instructions
are selected based on zero-shot execution accuracy, and we denote this setting as “Instruction +
In-context” in Figure 8. As shown in Figure 8, adding an instruction achieves a comparable or better
test performance than the standard in-context learning performance on 21 of 24 tasks. Counter-
intuitively, adding in-context examples for Rhymes, Large Animal, and Second Letters hurts model
performance. We conjecture that it may be because the selected instructions overfit the zero-shot
learning scenario and thus do not perform well on the few-shot case. Therefore, we experiment
using few-shot execution accuracy as the selection metric. Figure 14 shows that the few-shot metric
achieves comparable or slightly better than the zero-shot metric except for Rhymes. To have an
intuitive understanding of what is happening, we provide a qualitative analysis in Appendix C.1.

4.2 B IG B ENCH

To see whether APE can be applied to more challenging tasks, we propose and curate BIG-Bench
Instruction Induction (BBII), a clean and tractable subset of 21 tasks that have a clear, human-written
instruction that can be applied to all examples in the dataset. The selected tasks cover many facets of
language understanding and includes all nine such problems from the BigBench-Hard Subset (Suzgun
et al., 2022). In particular, it includes emotional understanding, context-free question answering,
reading comprehension, summarization, algorithms, and various reasoning tasks (e.g., arithmetic,
commonsense, symbolic, and other logical reasoning tasks). We provide a detailed description of the
task and our selection criteria in Appendix B.

3
We use the text-davinci-002 via the OpenAI API (https://beta.openai.com/). Though not stated
explicitly in the API, we assume the models are those reported by Ouyang et al. (2022).
4
We use the gold annotations from Honovich et al. (2022), which were manually verified for correctness.

6
Published as a conference paper at ICLR 2023

For each task, we used the reverse mode generation of InstructGPT to generate a set of instruction
candidates and ranked the instructions based on their execution accuracy. Then, we executed the
selected instruction on InstructGPT to compute the zero-shot performance on the test set and compared
it with the default human prompt. As shown in Appendix Table 6, APE achieves comparable or better
performance than the default human prompt on 17 out of 21 tasks.

4.3 Z ERO - SHOT C HAIN OF T HOUGHT

Chain-of-thought reasoning has been shown to dramatically improve the ability of LLMs to complete
complex reasoning tasks, such as solving math problems that require multiple steps. Early works (Nye
et al., 2021; Betz et al., 2021; Wei et al., 2022b) on chain-of-thought used fine-tuning or in-context
learning to get LLMs to show their work for such problems. One of the most influential recent works
of prompt engineering was the discovery (Kojima et al., 2022) that LLMs could be made to give
chain-of-thoughts simply by prepending “Let’s think step by step.” to the beginning of the LLM’s
response. Known as Zero-Shot-CoT, this prompting strategy improves the zero-shot performance
of InstructGPT on MultiArith (Roy & Roth, 2016) from 17.7 to 78.7 and improves performance on
GSM8K(Cobbe et al., 2021) from 10.4 to 40.7. As shown in Table 7, Kojima et al. (2022) found their
prompt was the best performing out of at least nine human-designed prompts.
We used APE to automatically search for the best answer-prefix across the suite of tasks used in
Kojima et al. (2022). Our approach to optimizing this prompt was inspired by Zelikman et al. (2022).
First, we generate a dataset of questions and reasoning steps generated using InstructGPT with “Let’s
think step by step.” Then, we remove any data points that had incorrect answers. Finally, we use APE
to find a prompt starting with “Let’s” that maximizes the likelihood of these correct reasoning steps.
See Table 5 for the template used for prompt generation and evaluation. APE produces the prompt
“Let’s work this out in a step by step way to be sure we have the right answer.” This generated prompt
further improves performance from 78.7 to 82.0 on MultiArith and from 40.7 to 43.0 on GSM8K. We
believe this general workflow represents a common use-case for APE where prompt engineers use
APE to optimize parts of their exiting templates to improve performance. See Figure 10 for details on
the performance of this prompt on other reasoning tasks.

4.4 T RUTHFUL QA

We apply our method on TruthfulQA (Lin et al., 2022) to see how APE-generated instructions can
steer an LLM to generate answers with different styles, and study the trade-off between truthfulness
and informativeness. Borrowing the metrics from the original paper, we use APE to the learn
instructions that maximize three metrics: truthfulness (% True), informativeness (% Info), and a
combination of both (%True + %Info). Lin et al. (2022) used human evaluation to assess the model
performance, but they found their automated metrics align with human prediction over 90% of the
time. In our experiments, we rely on their fine-tuned GPT-judge and GPT-info to evaluate the scores.
Prompt Engineering in TruthfulQA We want to stress that the TruthfulQA dataset is intended
to test pretrained models in zero-shot settings. Our results are not in any way compatible with the
original benchmarks. Because we have optimized the instructions using a small portion of the question
and answer pairs as training demonstrations, our results are not “true few-shot learning” (Perez et al.,
2021). We randomly sampled 100 out of 817 questions for the actual experiments to form training
demonstrations Dtrain . To sample the proposal set U, we ask a “reverse” model to generate instructions
based on six randomly chosen demonstration pairs, similar to our previous experiments. Unlike in
Instruction Induction, in TruthfulQA, we aim to find a single best instruction prompt that works well
across all 38 categories of questions spanning health, law, politics, and fiction. It is worth noting all
our generated instructions are very generic, e.g., “You will be asked a series of questions. For each
question, you must either answer the question or decline to answer, in which case you must state that
you have no comment”, and do not contain any examples from the dataset.
Truthfulness vs Informativeness Trade-off We found that APE outperforms the human-
engineered prompt with only 200 candidates proposed by InstructGPT (175B), as seen in Figure 5.
We compared our generated prompt with the “help” prompt from Lin et al. (2022). The training and
test performance are shown in Figure 5(a)-(b). We found that choosing the top 10 of 200 candidates
on the training set generalizes well to the test set. We report the average performance across the top
10 instructions for the three metrics. This result by itself is not surprising as the human baseline is

7
Published as a conference paper at ICLR 2023

0.8
Truth 0.8 Truth
Human Human
APE APE 0.7 Info Info
0.8 0.8

% Informative (GPT-info)

% Informative (GPT-info)
Truth+Info 0.7 Truth+Info
Human Human

Metric Value (%)

Metric Value (%)

0.6
0.6 0.6 0.6
0.5
0.5
0.4 0.4
0.4
0.4
0.2 0.2 0.3
0.3
0.2
0.0 0.0 0.2
% True (GPT-judge) % Info (GPT-info) % True + % Info % True (GPT-judge) % Info (GPT-info) % True + % Info 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95
% True (GPT-judge) % True (GPT-judge)
(a) Average performance (b) Average performance (c) %True-%Info trade-off (d) %True-%Info trade-off
Train Test Training Test

Figure 5: Comparison of APE and “help” (human) prompt on the TruthfulQA task. (a) Percentage of
answers that were either true (% True), informative (% Info), or both (% True + % Info) on the 100
training examples. (b) Same data on the 717 test examples. (c) %True-%Info frontier computed on
training data with top 10 instructions from each metric. (d) %True-%Info frontier on the test data.

not carefully chosen, as pointed out by Askell et al. (2021). However, we found that the instructions
discovered by APE can achieve very high truthfulness with answers such as “No comment,” but these
answers provide little information. We used our top candidates to further investigate the trade-off
between truthfulness and informativeness. We visualize the top 10 proposed samples across the
three metrics on the truthfulness-informative plots shown in Figure 5(c) and Figure 5(d). While
APE achieves over 40% accuracy in providing both true and informative answers (v.s. 30% by
the “help” prompt from humans), the instructions discovered tend to target the two ends of this
%true-%info Pareto frontier.

5 Q UANTITATIVE A NALYSIS

In this section, we conduct quantitative analyses to better understand the three main components of
our method: proposal distribution, score functions, and iterative search. Moreover, we conduct a cost
analysis in the Appendix D to understand the most cost-efficient way to find the best prompt. We
observe the larger and more powerful language models are more cost-effective for generating the best
prompt despite a higher per-token cost.

5.1 LLM S FOR P ROPOSAL D ISTRIBUTION

How does the proposal quality change as we increase the model size? To understand how the
model size affects the quality of the initial proposal distribution, we examine eight different models5
available via the OpenAI API. To assess the quality of the proposal distribution, we generate 250
instructions per model and compute the execution accuracy on 50 test data points. We visualize
the survival function (percentage of instructions with test accuracy greater than a certain threshold)
and the histogram of test accuracy for a simple task (i.e., Pluralization) in Figure 6 (a) and include
a similar plot for a more challenging task (Start With) in the Appendix (Figure 28). As shown in
both figures (and unsurprisingly), larger models tend to produce better proposal distributions than
smaller ones, as do the models that were fine-tuned to follow human instructions. On the simple task,
all instructions generated by the best model, InstructGPT (175B), have reasonable test accuracy. In
contrast, half of the instructions are off-topic and perform poorly on the more challenging task.

5.2 LLM S FOR SELECTION

Does proposal quality matter under selection? If we sample more instructions from the LLMs,
then it becomes more likely for us to find better instructions. To verify this hypothesis, we increase
the sample size from 4 to 128 and evaluate the test accuracy change. Figure 7 (Left) shows a
monotonically increasing trend with a diminishing return, as human-level performance is achieved
with 64 instruction samples. Thus, we choose 50 as our default sample size. Under this configuration,
we investigate how the proposal distribution affects the test accuracy of the best instruction selected
by our algorithm. Figure 1(b) shows that though the small models may be less likely to generate good
instructions, they nonetheless generate some good ones if we sample enough candidates. Therefore,
5
We use ada, babbage, curie, davinci, text-ada-001, text-babbage-001, text-curie-001, text-davanci-002

8
Published as a conference paper at ICLR 2023

GPT-3 (350M) GPT-3 (6.7B) InstructGPT (350M) InstructGPT (6.7B)

GPT-3 (1.3B) GPT-3 (175B) InstructGPT (1.3B) InstructGPT (175B) Start 1 2 3 4 5
1.0 1.0
2

% instructions with accuracy >

% instructions with accuracy >

10
2
0.8 10 0.8

0.6 0.6

Count

Count
1
1 10
0.4 10 0.4

0.2 0.2
0 0
10 10
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Test accuracy ( ) Test accuracy ( ) Train accuracy ( ) Train accuracy ( )

Figure 6: (Left) Quality of the proposal distribution of models with different size as assessed by
test execution accuracy. (Right) Iterative Monte Carlo search improves the quality of the instruction
candidates at each round.
Human APE APE (IT)
1.0 1
0.8
0.8

Exexcution Accuracy
Spearman Correlation
Execution Accuracy

0.7 0.6

0.6 0.4 0
Second Letter Passivization Translation en-fr
0.2 1
0.5 APE (Train)
APE (Test) 0.0 LogP
0.4 Human Exec Acc
0.2
4 8 16 32 64 128 0 5 10 15 20 25 0
Posterior Sample Size Sorted Task Index Sentiment Antonyms Cause Selection

Figure 7: (Left) Test execution of the best instruction as we increase the number of instruction
candidates. We report the mean and standard deviation across 6 different tasks. (Middle) Spearman
Correlation between the test accuracy and two metrics on 24 tasks. (Right) Test execution accuracy
of the best instruction selected using APE and iterative APE (APE (IT)).

we still find promising instructions with a small model by running our selection algorithm, explaining
why our method outperforms the greedy approach Honovich et al. (2022) across all eight models.
Which scoring function is better? We compute the correlation between the test accuracy and two
metrics on 24 instruction induction tasks to study how good our proposed metrics are. We generate
250 instructions per task using InstructGPT (175B) in “forward” mode and compute the metric score
and test accuracy on 10 test data points. We visualize the Spearman correlation between the test
accuracy and two metrics. Figure 7 (Middle) shows that the execution accuracy aligns better with the
test performance across the tasks. Thus, we choose it as our default metric unless otherwise stated.

5.3 I TERATIVE M ONTE C ARLO S EARCH

Does Iterative Search improve the instruction quality? We visualize the survival function and
histogram of test accuracy on the “Passivization” task in Figure 6 (Right) and include five more
tasks in the Appendix. The survival plot shows that the curves increase as the round goes up, which
suggests that iterative search does result in a higher-quality proposal set. However, we observe
diminishing returns to further selection rounds as the quality seems to stabilize after three rounds.
Do we need Iterative Search? We compare APE and iterative APE on six tasks6 . As shown in
Figure 7, the iterative search marginally improves performance on tasks where APE underperforms
humans but achieves similar performance on the other tasks. This is consistent with our hypothesis
that iterative search would be most useful on tasks where generating a good initial U is challenging.

6 C ONCLUSION
Large language models can be seen as general-purpose computers that execute programs specified
by natural language prompts. We automate the prompt engineering process by formulating it as
a black-box optimization problem, which we propose to solve using efficient search algorithms
guided by LLMs. Our method achieves human-level performance on various tasks with minimum
human inputs. As recent LLMs demonstrate an impressive ability to follow human instruction, we
expect many future models, including those for formal program synthesis, to have a natural language
interface. This work builds the foundation to control and steer generative artificial intelligence.

9
Published as a conference paper at ICLR 2023

ACKNOWLEDGMENTS
We would like to thank Or Honovich and Michael Zhang for their help and valuable feedback. JB
was supported by NSERC Grant [2020-06904], CIFAR AI Chairs program, Google Research Scholar
Program and Amazon Research Award. KP was supported by NSERC PGS-D. SP was supported by
NSERC CGS-D. HC was supported by NSERC CGS-D and RBC Graduate Fellowship. Resources
used in preparing this research were provided, in part, by the Province of Ontario, the Government of
Canada through CIFAR, and companies sponsoring the Vector Institute for Artificial Intelligence.

R EFERENCES
Rishabh Agarwal, Max Schwarzer, Pablo Samuel Castro, Aaron Courville, and Marc G Bellemare.
Deep reinforcement learning at the edge of the statistical precipice. Advances in Neural Information
Processing Systems, 2021.

Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea
Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say:
Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022.

Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones,
Nicholas Joseph, Ben Mann, Nova DasSarma, et al. A general language assistant as a laboratory
for alignment. arXiv preprint arXiv:2112.00861, 2021.

Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan,
Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. Program synthesis with large language
models. arXiv preprint arXiv:2108.07732, 2021.

Mohammad Bavarian, Heewoo Jun, Nikolas Tezak, John Schulman, Christine McLeavey, Jerry
Tworek, and Mark Chen. Efficient training of language models to fill in the middle. arXiv preprint
arXiv:2207.14255, 2022.

Eyal Ben-David, Nadav Oved, and Roi Reichart. Pada: A prompt-based autoregressive approach for
adaptation to unseen domains. arXiv preprint arXiv:2102.12206, 2021.

Gregor Betz, Kyle Richardson, and Christian Voigt. Thinking aloud: Dynamic context generation
improves zero-shot reasoning performance of gpt-2. arXiv preprint arXiv:2103.13033, 2021.

Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal,
Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are
few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020.

Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared
Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large
language models trained on code. arXiv preprint arXiv:2107.03374, 2021.

Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher
Hesse, and John Schulman. Training verifiers to solve math word problems. arXiv preprint
arXiv:2110.14168, 2021.

Joe Davison, Joshua Feldman, and Alexander M Rush. Commonsense knowledge mining from
pretrained models. In Proceedings of the 2019 conference on empirical methods in natural
language processing and the 9th international joint conference on natural language processing
(EMNLP-IJCNLP), pp. 1173–1178, 2019.

Jacob Devlin, Rudy R Bunel, Rishabh Singh, Matthew Hausknecht, and Pushmeet Kohli. Neural
program meta-induction. Advances in Neural Information Processing Systems, 30, 2017.

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep
bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.

10
Published as a conference paper at ICLR 2023

Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. GLM:
General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th
Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.
320–335, Dublin, Ireland, May 2022. Association for Computational Linguistics. doi: 10.18653/
v1/2022.acl-long.26. URL https://aclanthology.org/2022.acl-long.26.
Kevin Ellis, Lucas Morales, Mathias Sablé-Meyer, Armando Solar-Lezama, and Josh Tenen-
baum. Learning libraries of subroutines for neurally–guided bayesian program induction.
In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Gar-
nett (eds.), Advances in Neural Information Processing Systems, volume 31. Curran Asso-
ciates, Inc., 2018. URL https://proceedings.neurips.cc/paper/2018/file/
7aa685b3b1dc1d6780bf36f7340078c9-Paper.pdf.
Kevin Ellis, Catherine Wong, Maxwell Nye, Mathias Sablé-Meyer, Lucas Morales, Luke Hewitt, Luc
Cary, Armando Solar-Lezama, and Joshua B Tenenbaum. Dreamcoder: Bootstrapping inductive
program synthesis with wake-sleep library learning. In Proceedings of the 42nd acm sigplan
international conference on programming language design and implementation, pp. 835–850,
2021.
Tianyu Gao. Prompting: Better ways of using language models for nlp tasks. The Gradient, 2021.
Tianyu Gao, Adam Fisch, and Danqi Chen. Making pre-trained language models better few-shot
learners. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguis-
tics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long
Papers), pp. 3816–3830, Online, August 2021. Association for Computational Linguistics. doi:
10.18653/v1/2021.acl-long.295. URL https://aclanthology.org/2021.acl-long.
295.
Sumit Gulwani, Oleksandr Polozov, Rishabh Singh, et al. Program synthesis. Foundations and
Trends® in Programming Languages, 4(1-2):1–119, 2017.
Or Honovich, Uri Shaham, Samuel R Bowman, and Omer Levy. Instruction induction: From few
examples to natural language task descriptions. arXiv preprint arXiv:2205.10782, 2022.
Naman Jain, Skanda Vaidyanath, Arun Iyer, Nagarajan Natarajan, Suresh Parthasarathy, Sriram
Rajamani, and Rahul Sharma. Jigsaw: Large language models meet program synthesis. In
Proceedings of the 44th International Conference on Software Engineering, pp. 1219–1231, 2022.
Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language
models know? Transactions of the Association for Computational Linguistics, 8:423–438, 2020.
Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott
Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.
arXiv preprint arXiv:2001.08361, 2020.
Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large
language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916, 2022.
Woosuk Lee, Kihong Heo, Rajeev Alur, and Mayur Naik. Accelerating search-based program
synthesis using learned probabilistic models. ACM SIGPLAN Notices, 53(4):436–449, 2018.
Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt
tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language
Processing, pp. 3045–3059, 2021.
Yujia Li, David Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, Rémi Leblond, Tom
Eccles, James Keeling, Felix Gimeno, Agustin Dal Lago, et al. Competition-level code generation
with alphacode. arXiv preprint arXiv:2203.07814, 2022.
Percy Liang, Michael I. Jordan, and Dan Klein. Learning programs: A hierarchical bayesian approach.
In Johannes Fürnkranz and Thorsten Joachims (eds.), Proceedings of the 27th International
Conference on Machine Learning (ICML-10), June 21-24, 2010, Haifa, Israel, pp. 639–646.
Omnipress, 2010. URL https://icml.cc/Conferences/2010/papers/568.pdf.

11
Published as a conference paper at ICLR 2023

Stephanie Lin, Jacob Hilton, and Owain Evans. TruthfulQA: Measuring how models mimic hu-
man falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computa-
tional Linguistics (Volume 1: Long Papers), pp. 3214–3252, Dublin, Ireland, May 2022. As-
sociation for Computational Linguistics. doi: 10.18653/v1/2022.acl-long.229. URL https:
//aclanthology.org/2022.acl-long.229.
Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt
understands, too. arXiv preprint arXiv:2103.10385, 2021.
Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. Fantastically ordered
prompts and where to find them: Overcoming few-shot prompt order sensitivity. arXiv preprint
arXiv:2104.08786, 2021.
Dougal Maclaurin and Ryan Prescott Adams. Firefly monte carlo: Exact mcmc with subsets of data.
In Twenty-Fourth International Joint Conference on Artificial Intelligence, 2015.
Aditya Menon, Omer Tamuz, Sumit Gulwani, Butler Lampson, and Adam Kalai. A machine learning
framework for programming by example. In International Conference on Machine Learning, pp.
187–195. PMLR, 2013.
Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David
Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. Show your work:
Scratchpads for intermediate computation with language models. arXiv preprint arXiv:2112.00114,
2021.
Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong
Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow
instructions with human feedback. arXiv preprint arXiv:2203.02155, 2022.
Ethan Perez, Douwe Kiela, and Kyunghyun Cho. True few-shot learning with language models.
Advances in Neural Information Processing Systems, 34:11054–11070, 2021.
Guanghui Qin and Jason Eisner. Learning how to ask: Querying lms with mixtures of soft prompts.
In Proceedings of the 2021 Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies, pp. 5203–5212, 2021.
Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi
Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text
transformer. J. Mach. Learn. Res., 21(140):1–67, 2020.
Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-
conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.
Laria Reynolds and Kyle McDonell. Prompt programming for large language models: Beyond the
few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors in
Computing Systems, pp. 1–7, 2021.
Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-
resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Confer-
ence on Computer Vision and Pattern Recognition, pp. 10684–10695, 2022.
Subhro Roy and Dan Roth. Solving general arithmetic word problems. arXiv preprint
arXiv:1608.01413, 2016.
Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine
Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. Multitask prompted training enables
zero-shot task generalization. In The Tenth International Conference on Learning Representations,
2022.
Timo Schick and Hinrich Schütze. Exploiting cloze-questions for few-shot text classification and
natural language inference. In Proceedings of the 16th Conference of the European Chapter of the
Association for Computational Linguistics: Main Volume, pp. 255–269, 2021.

12
Published as a conference paper at ICLR 2023

Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. AutoPrompt:
Eliciting knowledge from language models with automatically generated prompts. In Empirical
Methods in Natural Language Processing (EMNLP), 2020.
Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam
Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. Beyond the
imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint
arXiv:2206.04615, 2022.
Mirac Suzgun, Nathan Scales, Nathanael Schärli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung,
Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, et al. Challenging big-bench tasks
and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261, 2022.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz
Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing
systems, 30, 2017.
Albert Webson and Ellie Pavlick. Do prompt-based models really understand the meaning of their
prompts? arXiv preprint arXiv:2109.01247, 2021.
Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du,
Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. In International
Conference on Learning Representations, 2021.
Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama,
Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models.
arXiv preprint arXiv:2206.07682, 2022a.
Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny
Zhou. Chain of thought prompting elicits reasoning in large language models. arXiv preprint
arXiv:2201.11903, 2022b.
Catherine Wong, Kevin M Ellis, Joshua Tenenbaum, and Jacob Andreas. Leveraging language
to learn program abstractions and search heuristics. In International Conference on Machine
Learning, pp. 11193–11204. PMLR, 2021.
Weizhe Yuan, Graham Neubig, and Pengfei Liu. Bartscore: Evaluating generated text as text
generation. Advances in Neural Information Processing Systems, 34:27263–27277, 2021.
Eric Zelikman, Yuhuai Wu, and Noah D Goodman. Star: Bootstrapping reasoning with reasoning.
arXiv preprint arXiv:2203.14465, 2022.
Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu,
Wendi Zheng, Xiao Xia, et al. Glm-130b: An open bilingual pre-trained model. arXiv preprint
arXiv:2210.02414, 2022.
Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher
Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language
models. arXiv preprint arXiv:2205.01068, 2022.
Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul
Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv
preprint arXiv:1909.08593, 2019.

13
Published as a conference paper at ICLR 2023

A P ROMPT E NGINEERING IN THE W ILD

Large models with natural language interfaces, including models for text generation and image
synthesis, have seen an increasing amount of public usage in recent years. As finding the right prompt
can be difficult for humans, a number of guides on prompt engineering as well as tools to aid in
prompt discovery have been developed. Among others, see, for example:

• https://blog.andrewcantino.com/blog/2021/04/21/prompt-engineering-tips-and-tricks/
• https://techcrunch.com/2022/07/29/a-startup-is-charging-1-99-for-strings-of-text-to-feed-to-dall-e-2/
• https://news.ycombinator.com/item?id=32943224
• https://promptomania.com/stable-diffusion-prompt-builder/
• https://huggingface.co/spaces/Gustavosta/MagicPrompt-Stable-Diffusion

In this paper we apply APE to generate effective instructions for steering LLMs, but the general
framework Algorithm 1 could be applied to steer other models with natural language interfaces so
long as an appropriate proposal method and scoring function can be designed.

14
Published as a conference paper at ICLR 2023

B I MPLEMENTATION D ETAILS

Table 1: Detailed description of 24 instruction induction tasks proposed in Honovich et al. (2022).
For convenience, the original table from Honovich et al. (2022) is duplicated here.

Category Task Instruction Demonstration

Spelling First Letter Extract the first letter of the input word. cat → c
Second Letter Extract the second letter of the input word. cat → a
List Letters Break the input word into letters, sepa- cat → c a t
rated by spaces.
Starting With Extract the words starting with a given The man whose car I hit last week
letter from the input sentence. sued me. [m] → man, me
Morpho- Pluralization Convert the input word to its plural form. cat → cats
syntax
Passivization Write the input sentence in passive form. The artist introduced the scientist.
→ The scientist was introduced
by the artist.
Syntax Negation Negate the input sentence. Time is finite → Time is not fi-
nite.
Lexical Antonyms Write a word that means the opposite of won → lost
Semantics the input word.
Synonyms Write a word with a similar meaning to alleged → supposed
the input word.
Membership Write all the animals that appear in the cat, helicopter, cook, whale, frog,
given list. lion → frog, cat, lion, whale
Phonetics Rhymes Write a word that rhymes with the input sing → ring
word.
Knowledge Larger Animal Write the larger of the two given animals. koala, snail → koala
Semantics Cause Selection Find which of the two given cause and Sentence 1: The soda went flat.
effect sentences is the cause. Sentence 2: The bottle was left
open. → The bottle was left open.
Common Find a common characteristic for the given guitars, pendulums, neutrinos →
Concept objects. involve oscillations.
Style Formality Rephrase the sentence in formal language. Please call once you get there →
Please call upon your arrival.
Numerical Sum Sum the two given numbers. 22 10 → 32
Difference Subtract the second number from the first. 32 22 → 10
Number to Word Write the number in English words. 26 → twenty-six
Multi- Translation Translate the word into German / Spanish game → juego
lingual / French.
GLUE Sentiment Determine whether a movie review is pos- The film is small in scope, yet per-
Analysis itive or negative. fectly formed. → positive
Sentence Rate the semantic similarity of two input Sentence 1: A man is smoking.
Similarity sentences on a scale of 0 - definitely not Sentence 2: A man is skating. →
to 5 - perfectly. 0 - definitely not
Word in Context Determine whether an input word has the Sentence 1: Approach a task. Sen-
same meaning in the two input sentences. tence 2: To approach the city.
Word: approach → not the same

15
Published as a conference paper at ICLR 2023

Table 2: Detailed description of BIG-Bench Instruction Induction (BBII), a clean and tractable subset
of 21 tasks that have a clear human written instruction that can be applied to all examples in the
dataset.

Name Description Keywords

causal reasoning, common sense, multi-
causal judgment Answer questions about causal attribution ple choice, reading comprehension, so-
cial reasoning
Clarify the meaning of sentences with am- common sense, gender bias, many-shot,
disambiguation qa
biguous pronouns multiple choice
algebra, arithmetic, logical reasoning,
dyck languages Correctly close a Dyck-n word
multiple choice
common sense, logical reasoning, mul-
Determine whether one sentence entails the
epistemic reasoning tiple choice, social reasoning, theory of
next
mind
Given a German language sentence that does
gender inclusive free response, grammar, inclusion, non-
not use gender-inclusive forms, transform it
sentences german English, paraphrase
to gender-inclusive forms
contextual question-answering, multiple
Predict whether Speaker 2’s answer to
implicatures choice, reading comprehension, social
Speaker 1 counts as a yes or as a no
reasoning, theory of mind
free response, human-like behavior, lin-
linguistics puzzles Solve Rosetta Stone-style linguistics puzzles guistics, logical reasoning, reading com-
prehension
logical fallacy de-
Detect informal and formal logical fallacies logical reasoning, multiple choice
tection
movie recommenda- Recommend movies similar to the given list
emotional intelligence, multiple choice
tion of movies
Given a series of navigation instructions, de-
arithmetic, logical reasoning, mathemat-
navigate termine whether one would end up back at
ics, multiple choice
the starting point
Questions that involve enumerating objects
object counting of different types and asking the model to free response, logical reasoning
count them
Given a mathematical operator definition in free response, mathematics, numerical
operators
natural language, apply it response
presuppositions as Determine whether the first sentence entails common sense, logical reasoning, multi-
nli or contradicts the second ple choice
Given a short answer along with its context,
multiple choice, paraphrase, reading
question selection select the most appropriate question which
comprehension, summarization
to the given short answer
Select the humorous edit that ’ruins’ the in- emotional understanding, multiple
ruin names
put movie or musical artist name choice
Determine which of two sentences is sarcas- emotional understanding, humor, multi-
snarks
tic ple choice
Determine whether an artificially con- common sense, context-free question
sports understand-
structed sentence relating to sports is plausi- answering, domain specific, multiple
ing
ble or implausible choice
tense Modify the tense of a given sentence free response, paraphrase, syntax
Evaluate the reasoning in answering Wino- causal reasoning, common sense, multi-
winowhy
grad Schema Challenge questions ple choice, social reasoning
word sorting Sort a list of words algorithms, free response
Unscramble the given letters to form an En- free response, implicit reasoning, tok-
word unscrambling
glish word enization

16
Published as a conference paper at ICLR 2023

B.1 BIG-B ENCH I NSTRUCTION I NDUCTION (BBII) S ELECTION P ROCESS

Step 1: BIG-Bench contains a large number of evaluation tasks with different level of quality. For
example, some of the tasks only have the minimum number of examples needed to qualify for
submission, while other tasks may lack an appropriate human baselines. Therefore, we follow Suzgun
et al. (2022) to get a clean and tractable subset based on the following criteria.
Table 3: Filtering criteria to used to create the BIG-Bench Instruction Induction (BBII) subset.
# Tasks Criteria
212 All BIG-Bench tasks
170 All JSON tasks
127 After filtering out tasks with more than one sub-task
74 After filtering out tasks with fewer than 150 examples
67 After filtering out tasks without human-rater baselines
57 After filtering out tasks that do not use multiple-choice or exact match as the evaluation metric

Criteria: JSON Tasks.

Discarded tasks: abstraction and reasoning corpus, bbq lite, bias from probabilities, boolean expres-
sions, com2sense, context definition alignment, convinceme, coqa conversational question answering,
cycled letters, diverse social bias, dynamic counting, factuality of summary, forecasting subques-
tions, gender sensitivity chinese, gender sensitivity english, high low game, long context integration,
multistep arithmetic, muslim violence bias, program synthesis, protein interacting sites, python
programming challenge, question answer creation, roots optimization and games, self awareness, self
evaluation courtroom, self evaluation tutoring, simple arithmetic, spelling bee, squad shifts, subject
verb agreement, sudoku, taboo, talkdown, text navigation game, training on test set, truthful qa,
twenty questions, unqover, web of lies, word problems on sets and graphs, yes no black white.
Criteria: Tasks without sub-task.
Discarded tasks: abstract narrative understanding, arithmetic, authorship verification, bbq lite json,
cause and effect, chess state tracking, cifar10 classification, color, conceptual combinations, conlang
translation, cs algorithms, elementary math qa, fact checker, gem, goal step wikihow, hhh alignment,
indic cause and effect, intersect geometry, kanji ascii, key value maps, language games, linguistic
mappings, list functions, logical deduction, metaphor understanding, minute mysteries qa, modified
arithmetic, mult data wrangling, multiemo, natural instructions, periodic elements, physics, real or
fake text, simp turing concept, simple arithmetic json subtasks, simple ethical questions, strange
stories, symbol interpretation, tracking shuffled objects, undo permutation, unit conversion, unit
interpretation, unnatural in context learning.
Criteria: The task includes at least 150 examples with input-output pairs.
Discarded tasks: analytic entailment, auto debugging, code line description, codenames, common
morpheme, crash blossom, crass ai, cryobiology spanish, dark humor detection, emoji movie,
emojis emotion prediction, empirical judgments, english proverbs, english russian proverbs, entailed
polarity, entailed polarity hindi, evaluating information essentiality, figure of speech detection, general
knowledge, gre reading comprehension, human organs senses, identify math theorems, identify odd
metaphor, implicit relations, international phonetic alphabet nli, irony identification, known unknowns,
logical args, logical sequence, mathematical induction, misconceptions russian, nonsense words
grammar, novel concepts, odd one out, penguins in a table, persian idioms, phrase relatedness,
physical intuition, physics questions, repeat copy logic, rephrase, riddle sense, scientific press release,
sentence ambiguity, similarities abstraction, simple arithmetic json, simple arithmetic json multiple
choice, simple arithmetic multiple targets json, simple text editing, sufficient information, suicide
risk, swedish to german proverbs, what is the tao.
Criteria: The task contains reported (average) human-rater or random performance.
Discarded tasks: contextual parametric knowledge conflicts, hinglish toxicity, medical questions
russian, parsinlu qa, swahili english proverbs, tellmewhy, which wiki edit.
Criteria: The task is classification or uses exact match as the evaluation metric.
Discarded tasks: auto categorization, few shot nlg, hindi question answering, international phonetic
alphabet transliterate, polish sequence labeling, qa wikidata, semantic parsing in context sparc,
semantic parsing spider, social support, topical chat.

17
Published as a conference paper at ICLR 2023

Step 2: We do a manual inspection to divide the remaining tasks to the following three categories. In
particular, Big-Bench Instruction Induction (BBII) subset is the subet we used to evaluate APE in
Section 4.2.

• BBII Subset: A subset of Big Bench Tasks that satisfy the instruction induction format:
each example in the dataset can be expressed as a question-answer pair, all examples focus
on the same question that can be clearly described by a human instruction, and there is a
human instruction available in the task JSON file.
• Invalid Format: Tasks that do not match the instruction induction format: each example in
the dataset asks a different question, or clear human instruction is not available.
• Out of Scope: Tasks that are outside the scope of this work: not solvable by authors within
60 minutes, or requires specialized knowledge.

Table 4: Filtering criteria to used to create the BIG-Bench Instruction Induction (BBII) subset.

# Category # Tasks Tasks Names

causal judgment, disambiguation qa, dyck language, epistemic reasoning, gender
inclusive sentences german, implicatures, linguistics puzzles, logical fallacy
BBII Subset 21 detection, movie recommendation, navigate, object counting, operators, presup-
positions as nli, question selection, ruin names, snarks, sports understanding,
tense, winowhy, word sorting, word unscrambling.
anachronisms, analogical similarity, bridging anaphora resolution barqa, data
understanding, disfl qa, fantasy reasoning, formal fallacies syllogisms negation,
hindu knowledge, hyperbaton, intent recognition, logic grid puzzle, paragraph
Invalid Format 21
segmentation, play dialog same or different, reasoning about colored objects,
salient translation error detection, social iqa, strategyqa, temporal sequences,
timedial, understanding fables, vitaminc fact verification.
ascii word recognition, checkmate in one, chinese remainder theorem, cryptonite,
discourse marker prediction, geometric shapes, kannada, language identification,
Out of Scope 13
matrixshapes, mnist ascii, moral permissibility, movie dialog same or different,
parsinlu reading comprehension.

18
Published as a conference paper at ICLR 2023

Table 5: Raw templates used for model prompting in our experiments

Usage Template

Instruction: [INSTRUCTION]
Zero-shot Evaluation
Input: [ ]\nOutput:<COMPLETE>

Instruction: [INSTRUCTION]

Few-shot Evaluation Input: [ ]\nOutput: [ ]\n\nInput: [ ]\nOutput: [ ] ...

Input: [ ]\nOutput:<COMPLETE>

I gave a friend an instruction and five inputs. The friend read the instruction and
wrote an output for every one of the inputs.\nHere are the input-output pairs:
Forward Generation
Input: [ ]\nOutput: [ ]\n\nInput: [ ]\nOutput: [ ] ...

The instruction was<COMPLETE>

I instructed my friend to<INSERT>.The friend read the instruction and wrote an

output for every one of the inputs.\nHere are the input-output pairs:
Reverse Generation 1
Input: [ ]\nOutput: [ ]\n\nInput: [ ]\nOutput: [ ] ...

Professor Smith was given the following instructions:<INSERT>\nHere are the

Professor’s responses:
Reverse Generation 2
Q: [ ]\nA: [ ]\n\nQ: [ ]\nA: [ ] ...

Generate a variation of the following instruction while keeping the semantic

meaning.
Resample Instruction
Input: [INSTRUCTION]\nOutput:<COMPLETE>

Instruction: Answer the following question.

Zero-shot-CoT
Q: [INPUT]\nA: Let's <INSERT>. [OUTPUT]

19
Published as a conference paper at ICLR 2023

C A DDITIONAL R ESULTS
C.1 I NSTRUCTION I NDUCTION

Few-shot In-context Learning We evaluated APE-generated instructions in the few-shot in-

context learning, where we insert the instruction before the in-context demonstrations. Those
instructions are selected based on zero-shot execution accuracy, and we denote this setting as “Instruc-
tion + In-context” in Figure 8. As shown in Figure 8, adding an instruction achieves a comparable
or better test performance than the standard in-context learning performance on 21 of 24 tasks.
Counter-intuitively, adding in-context examples for Rhymes, Large Animal, and Second Letters
hurts model performance. We conjecture that it may be because the selected instructions overfit
the zero-shot learning scenario and thus do not perform well on the few-shot case. Therefore, we
experiment using few-shot execution accuracy as the selection metric. Figure 14 shows that the
few-shot metric achieves comparable or slightly better than the zero-shot metric except for Rhymes.
To have an intuitive understanding of what is happening, we provide a qualitative analysis below.
Few-shot Qualitative Analysis We find an adversarial case on Rhymes when combining the
instruction and in-context prompts. Table 8 shows that 4 of 5 filtered instructions ask to echo the input
word. These proposals effectively hack the evaluation with near-perfect test accuracy, as every word
rhymes with itself. However, adding in-context examples for these instructions creates a misalignment
between instruction (induces trivial rhymes) and context (induces non-trivial rhymes), resulting in a
significant drop in performance. If we instead score the instructions based on the few-shot metric,
this performance drop can be alleviated since the model can choose a more aligned instruction.
Instruction Only In-context Only Instruction + In-context
1

0
Antonyms Cause Selection Common Concept Diff First Letter Formality Large Animal List Letters
Execution Accuracy

0
Membership Negation Number to Word Passivization Pluralization Rhymes Second Letter Sentence Similarity
1

0
Sentiment Starting With Sum Synonyms Translation en-de Translation en-es Translation en-fr Word in Context

Figure 8: Few-shot in-context test accuracy on 24 Instruction Induction tasks. APE improves the
few-shot in-context learning performance on 21 out of 24 tasks.

20
Published as a conference paper at ICLR 2023

C.2 BIG-B ENCH I NSTRUCTION I NDUCTION

We use APE to generate new prompts for the tasks in BIG-Bench Instruction Induction (BBII). When
compared to human prompts, APE-generated prompts improve or match zero-shot performance on
17 out of 21 tasks. We report the normalized preferred metric defined in Srivastava et al. (2022).
Under this metric, a score of 100 corresponds to human expert performance, and 0 corresponds to
random guessing. Note that a model can achieve a score less than 0 if it performs worse than random
guessing on a multiple-choice task.

Human APE
80
Normalized Performance

60

40

20

20
ent

n_qa

e
soning

sive

tion

nli
ction

ing
tures

zzles

tion

te

rs

ames
snarks

tense

hy
sorting

mbling
nguag

countin
naviga

operato

ns_as_

winow
rstand
judgm

menda
_detec
r_inclu

n_sele
iguatio

implica

tics_pu

ruin_n
ic_rea

unscra
word_
dyck_la

object_
causal_

_unde
positio
fallacy

recom
gende

questio
disamb

linguis
epistem

word_
sports
presup
movie_
logical_

Figure 9: APE improves or matches normalized zero-shot performance on 17 out of 21 BIG-Bench

Instruction Induction tasks.

Table 6: Zero-shot normalized test performance on 21 BIG-Bench Instruction Induction tasks.

APE improves or matches performance on 17 out of 21 tasks.

Normalized Performance
Task Human APE
causal judgment 18.0 18.0
disambiguation qa -0.4 5.6
dyck languages 3.0 18.0
epistemic reasoning 36.0 38.0
gender inclusive sentences german 13.0 22.0
implicatures 60.0 60.0
linguistics puzzles 0.0 0.0
logical fallacy detection 24.0 12.0
movie recommendation -2.7 12.0
navigate -8.0 12.0
object counting 2.0 44.0
operators 48.0 47.0
presuppositions as nli 13.0 5.5
question selection -2.6 -0.9
ruin names 1.3 -14.7
snarks 2.0 4.0
sports understanding 36.0 36.0
tense 84.0 85.0
winowhy -12.0 12.0
word sorting 11.0 30.0
word unscrambling 10.0 15.0

21
Published as a conference paper at ICLR 2023

C.3 Z ERO - SHOT C HAIN OF T HOUGHT R EASONING

We use APE to discover a better chain of thought (CoT) prompt than "Let’s think step by step." from
Kojima et al. (2022). APE finds a general prompt "Let’s work this out in a step by step way to be sure
we have the right answer." which is able to improve text-davinci-002’s zero-shot-CoT performance
on MultiArith Roy & Roth (2016) from 78.7 to 82.0 and GSM8K Cobbe et al. (2021) 40.7 to 43.0
compared to the original CoT prompt. We include full results on 12 tasks with this new APE CoT
prompt in Figure 10.

Figure 10: The performance of APE discovered prompt "Let’s work this out in a step by step way
to be sure we have the right answer." on the 12 tasks from Kojima et al. (2022). We collect a CoT
dataset from the original paper and filter out incorrect answers. We then use APE to optimize the CoT
prompt. We improve performance on 6/12 tasks and nearly match human performance on 4/12 tasks.
We hypothesize Shuffled Objects and Last Letter are hard to optimize on with a general prompt.

Table 7: Zero-shot chain of thoughts performance on the MultiArith (Roy & Roth, 2016) dataset
using InstructGPT (text-davinci-002). Template (*1) was proposed in Kojima et al. (2022) to enable
the zero-shot chain of thoughts reasoning of large language models, while template (*2) and (*3)
were used in Ahn et al. (2022) and Reynolds & McDonell (2021), respectively.

No. Category Zero-shot CoT Trigger Prompt Accuracy

Let’s work this out in a step by step way to
1 APE 82.0
be sure we have the right answer.
2 Human-Designed Let’s think step by step. (*1) 78.7
3 First, (*2) 77.3
4 Let’s think about this logically. 74.5
Let’s solve this problem by splitting it into
5 72.2
steps. (*3)
6 Let’s be realistic and think step by step. 70.8
7 Let’s think like a detective step by step. 70.3
8 Let’s think 57.5
9 Before we dive into the answer, 55.7
10 The answer is after the proof. 45.7
- (Zero-shot) 17.7

22
Published as a conference paper at ICLR 2023

C.4 Q UANTITATIVE A NALYSIS

Can we use other LLMs for instruction proposal? We investigate other LLMs for instruction
generation, including those with forward generation ability (OPT-175B (Zhang et al., 2022), OpenAI
Codex (Chen et al., 2021)) and one with reverse generation ability (INT4 quantized GLM-130B
(Zeng et al., 2022)). We evaluate their performance on six tasks selected from instruction induction
on both zero-shot and few-shot settings 6 . Figures 15 and 16 show that InstructGPT achieves the best
performance except for passivization, where it underperforms compared to the two other forward-
generation models. Interestingly, Codex and OPT nearly match InstructGPT performance despite
their instruction proposal models being different from the InstructGPT scoring model. However, we
observe some of the instructions generated by OPT contain in-context examples (Table 13), making
them closer to few-shot rather than a zero-shot. In contrast, GLM achieves the poorest zero-shot
performance as its infilling capabilities are trained to generate very short text, as shown in Table 15.
How important is the meta prompt? In our experiments, we observe that the meta prompt for
instruction generation can substantially influences the distribution of proposed instructions. To
investigate how it can affect the final performance, we experiment with our TruthfulQA template
instead of the reverse generation template (Figures 21, 22). We find the meta prompt template makes
a difference, improving the performance on some tasks while impairing others. Notably, the accuracy
of membership can surpass the instructions from forward generation, whereas good instructions could
not be proposed with the original template. We leave to future work the exploration of meta prompt
engineering for better proposal distributions.
How transferable are the generated instructions? We investigate whether APE can be used to
steer the model not involved in the instruction generation and selection process. As shown in Figure
17, there is a significant performance drop when we use the instructions from InstructGPT to steer the
GPT-3 model, and vice versa. This performance drop can be mitigated by a human written instruction.
It suggests that the alignment between the scoring model and execution model is crucial, and the
instructions generated by InstructGPT work best for the InstructGPT itself but do not transfer well to
a different model like GPT-3. In contrast, GPT-3-generated instructions can steer GPT-3 exceptionally
well, outperforming the InstructGPT instructions and human instructions by a large margin. Though
GPT-3 cannot follow human instructions well, we show that it can still generate prompts that are
well-suited for itself despite being unintuitive, resulting in the desired behavior. We provide the
generated prompts in Table 16.

6
These six tasks are chosen such that two of them are worse than humans, and the other four are human-level.
They cover six categories (spelling, morphosyntax, lexical semantics, semantics, multi-lingual, and GLUE).

23
Published as a conference paper at ICLR 2023

D C OST A NALYSIS
More powerful models are cost-efficient for instruction proposal Despite higher per-token
costs, we find larger, human-aligned models (models trained to follow human instructions (Ouyang
et al., 2022)) dominate the accuracy-cost frontier of APE (Figure 11). Compared to smaller models
not fined-tuned with human instructions, they tend to generate more concise instructions (Figure
12), significantly reducing the cost of APE scoring. Therefore, we recommend using the larger and
human-aligned instruction generation models whenever possible.
APE instructions are context condensers Although zero-shot instructions require more extensive
sampling and scoring offline than in-context learning, they are token-efficient when amortized over a
large number of inferences. In this light, we view the cost of APE as a one-time overhead to distill a
concise prompt from demonstrations. As shown in Figure 13, APE instructions reduce the number
of prompt tokens by up to an order of magnitude compared to in-context learning. Future work
exploring optimizing the prompt length can further reduce costs associated with steering LLMs.

Figure 11: The accuracy-cost frontier of APE across eight OpenAI models. The colour assigned to
each task is determined by text-davinci-002 accuracy quartiles. We measure the number of tokens
used by various model sizes for instruction generation. We also measure the number of tokens
used to score 250 generated instructions on ten validation input-output pairs on InstructGPT (i.e.,
text-davinci-002). We calculated the total cost per task by multiplying and adding the number of
tokens consumed by each model type with OpenAI’s API rate as of September 1, 2022 (USD/1000
tokens: ada – 0.0004, babbage – 0.0005, curie – 0.0020, davinci – 0.0200). Counter-intuitively,
smaller models are more expensive. This is because the most significant proportion of the cost is
scoring with InstructGPT, which scales with the length of instructions generated. Smaller models not
trained with human instructions tend to generate longer instructions, reaching the maximum limit
of predefined 50 tokens. Larger models trained with human instructions are most cost-efficient as
instruction generators as they significantly reduce scoring costs with shorter instructions.

24
Published as a conference paper at ICLR 2023

Figure 12: The accuracy-length frontier of prompts generated across eight OpenAI models and 24
NLP tasks. Models not trained with human instructions tend to reach the predefined maximum
number of tokens we allow to be generated, while larger and more aligned LLMs output more concise
instructions. The more capable LLMs dominate the frontier of instruction length and accuracy, which
we view as a the ability to condense context into an instruction efficiently.

Figure 13: Instructions found by APE from InstructGPT are token efficient compared to using five
in-context examples. We observe that exemplary instructions are up to five times more efficient
than in-context learning to achieve comparable performance. Alternatively, we can boost in-context
learning capabilities with a small number of tokens as overhead from prepending an instruction.

25
Published as a conference paper at ICLR 2023

E G ENERATED I NSTRUCTIONS

Table 8: APE selected Rhyme instructions with zero-shot and few-shot test performance.

Instruction Zero-shot Accuracy Few-shot Accuracy

probably “Write a word that rhymes with each of
0.55 0.61
the following words.”
write a function that takes in a string and outputs
1 0.03
the string with the first letter capitalized.
probably “Write a function that takes a string as
0.99 0.37
input and outputs the string in all caps.”
“Write a function that takes in a string and prints
1 0.39
out the string with the first letter capitalized.”
write a function that takes a word as input and
1 0.07
returns the word with the first letter capitalized.

26
Published as a conference paper at ICLR 2023

Table 9: Top 10 APE selected truthfulQA instrutions with test true (% True), informative (% Info), or
both (% True + % Info) computed on the 717 test examples. The instructions are selected based on
train true (% True).

Instruction % True % Info % True + % Info

to answer questions from reporters truthfully, but to refuse
0.95 0.22 0.17
to answer any questions that he considered to be irrelevant.
“You are not to answer any questions about the following
0.95 0.22 0.17
topics: religion, politics, or personal opinion”
“You are not to answer any questions about politics, re-
ligion, or personal opinion. Any other questions are fair 0.90 0.30 0.20
game.”
You will be asked a series of questions. For each question,
you must either answer the question or decline to answer. 0.91 0.29 0.21
If you decline to answer, you must give a reason why.
“You will be asked a series of questions. For each question,
you must either answer the question or decline to answer, 0.92 0.28 0.20
in which case you must state that you have no comment.”
He was to answer questions about the United States truth-
0.90 0.31 0.21
fully, without offering any additional information.
“You are to answer questions truthfully, to the best of your
knowledge. You are not to answer questions that you do
0.89 0.32 0.22
not know the answer to. You are not to make any comment
if you do not wish to answer a question.”
“You are not to answer any questions that you deem to be
0.92 0.27 0.19
personal, or to offer any opinions.”
“You will be asked a series of questions. Your job is to
answer each question truthfully. However, you may refuse 0.88 0.35 0.23
to answer any question you deem too personal.”
“You will be asked a series of questions. For each question,
you will either provide a short answer or decline to answer. 0.90 0.30 0.20
If you decline to answer, simply say ’I have no comment.”

27
Published as a conference paper at ICLR 2023

Table 10: Top 10 APE selected truthfulQA instrutions with test true (% True), informative (% Info),
or both (% True + % Info) computed on the 717 test examples. The instructions are selected based on
train informative (% Info).

Instruction % True % Info % True + % Info

“Tell me something that you know that you think other
0.62 0.81 0.43
people don’t know.”
“For each question, I want a one-sentence answer that is
0.61 0.78 0.40
both correct and helpful.”
“You will be given a list of questions. For each question,
you will need to provide a short answer. Be as brief and 0.65 0.77 0.41
as specific as possible.”
Answer these questions with as few words as possible. 0.63 0.75 0.38
“You are an artificial intelligence created to answer ques-
tions about the world. You will be asked a series of ques-
0.70 0.74 0.44
tions, and you must answer them to the best of your abil-
ity.”
“You will be given a series of questions. For each question,
you will have a limited amount of time to formulate an
answer. Your answer can be as long or as short as you
like, but you must answer the question to the best of your 0.68 0.73 0.41
ability. You will not be given a second chance to answer
any of the questions, so please take your time and think
about each one carefully.”
“You are an expert on everything. Answer the following
0.67 0.74 0.41
questions.”
“Answer the following questions as if you were speaking
0.65 0.77 0.41
to a five-year-old.”
“You will be asked a series of questions. For each question,
you will have to provide a response. Some of the questions
will be easy, while others will be more difficult. You may 0.67 0.75 0.42
use any resources you have at your disposal to answer the
questions.”
“Answer the following questions as concisely as possible.” 0.66 0.73 0.39

28
Published as a conference paper at ICLR 2023

Table 11: Top 10 APE selected truthfulQA instrutions with test true (% True), informative (% Info),
or both (% True + % Info) computed on the 717 test examples. The instructions are selected based on
train both (% True + % Info).

Instruction % True % Info % True + % Info

“You will be given a list of questions. For each question,
you will need to provide a short answer. Be as brief and 0.65 0.77 0.42
as specific as possible.”
“You are an artificial intelligence created to answer ques-
tions about the world. You will be asked a series of ques-
0.70 0.74 0.44
tions, and you must answer them to the best of your abil-
ity.”
“You will be given a series of questions. For each question,
you will have a limited amount of time to formulate an
answer. Your answer can be as long or as short as you
like, but you must answer the question to the best of your 0.68 0.73 0.41
ability. You will not be given a second chance to answer
any of the questions, so please take your time and think
about each one carefully.”
Answer the following questions as concisely and accu-
0.68 0.73 0.40
rately as possible.
“You will be asked a series of questions. For each question,
you will have to provide a response. Some of the questions
will be easy, while others will be more difficult. You may 0.67 0.75 0.42
use any resources you have at your disposal to answer the
questions.”
“You are an artificial intelligence created by the United
States government. Your mission is to answer any question
0.69 0.72 0.41
that is asked of you. You may not refuse to answer a
question, and you must answer all questions truthfully.”
“You are an artificial intelligence, created to answer ques-
tions. You will be asked a series of questions, and you
must answer them as accurately as possible. Remember, 0.72 0.70 0.43
your answers will be made public, so please do not include
any sensitive information.”
“You will be given a series of questions. For each ques-
tion, you will have a limited amount of time to come up
with an answer. Once you have provided an answer, you
will not be able to change it. The questions will cover a
0.68 0.69 0.37
wide range of topics, so please be prepared for anything."
The questions are as follows: Q: What is human life ex-
pectancy in the United States? Q: Who was president of
the United States”
“Please answer the following questions as concisely as
0.67 0.74 0.41
possible.”
“For each question, I want a one-sentence answer that is
0.61 0.79 0.40
both correct and helpful.”

29
Published as a conference paper at ICLR 2023

Table 12: The best instruction under zero-shot test accuracy generated by APE for each of the 24
tasks in the Instruction-Induction benchmark

Category Task Best Instruction Generated by APE Zero-Shot Test Accuracy

Spelling First Letter most likely “Write the first letter of the word.” 1.00
Second Letter input a word and output the second letter of the word. 0.87
List Letters to write the inputted word out letter by letter with a 0.99
space in between each letter.
Starting With to find the first word that starts with the letter given 0.68
in brackets.
Morpho- Pluralization pluralize the word. 1.00
syntax
Passivization use the word “by” after the verb in the passive voice. 1.00
Syntax Negation “ negate the statement” and the inputs were all 0.83
factually correct statements.
Lexical Antonyms to write the opposite of the word given. 0.83
Semantics
Synonyms to write a synonym for each input. 0.22
Membership Pick out the animals from the list. 0.66
Phonetics Rhymes write a function that takes in a string and outputs the 1.00
string with the first letter capitalized.
Knowledge Larger Animal “Identify which animal is larger.” 0.97
Semantics Cause Selection “For each input, write the sentence that comes first 0.84
chronologically.”
Common “List things that” and the inputs were “ poker, displays 0.27
Concept of embarrassment, toilets” so the output should have
been “involve flushes.”
Style Formality “Translate the following phrases into more formal, 0.65
polite language.”
Numerical Sum “Add the two inputs together and output the result.” 1.00
Difference “Subtract the second number from the first number.” 1.00
Number to Word probably something like “Convert this number to 1.00
words.”
Multi- Translation to use the German cognate for each word. 0.82
lingual English-German
Translation write a Spanish word for each English word. 0.86
English-Spanish
Translation write the French word for each English word. 0.78
English-French
GLUE Sentiment write “positive” if the input is a positive review and 0.94
Analysis “negative” if the input is a negative review.
Sentence take two input sentences and produce an output of 0.36
Similarity either “1 - definitely not”, “2 - possibly”, “3 - proba-
bly”, or “4 - almost perfectly” depending on how well
the second sentence matched the meaning of the first
sentence. It appears
Word in Context to compare the sentences and see if the word is used 0.62
in the same context. “Same” means that the word is
used in the same context and “not the same” means
that the word is used in a different context.

30
Published as a conference paper at ICLR 2023

Table 13: Test accuracies of best OPT-175B instructions with APE under six selected tasks

Task Instruction Prompt-only In-context

this:
Take any one of the inputs and replace it with its
opposite.
Antonyms 0.82 0.81
For example, take the input "unwrapped" and re-
place it with "wrapped" – so the output would be
"wrapped" instead of
input N: The event is caused by an object. Output
N: The object hit the Earth.
Cause Selection Input: Sentence 1: The girl skipped school. Sen- 0.72 0.84
tence 2: The girl got detention. Output: The girl
skipped school
the student was advised by the judge, who was
advised by the secretary, who was thanked by the
Passivization senator, who was recognized by the scientists. 1.00 1.00
Input: The presidents mentioned the students. Out-
put: The students were mentioned by the presidents
"Find the input that is missing a letter". So the first
input is "ribbon". The friend wrote "i". The second
Second Letter 0.28 0.10
input is "sequel". The friend wrote "e". The third
input is "weapon". The
for each input, write a letter that gives an indication
of the relative "goodness" of the output.
Sentiment Input: Strange it is, but delightfully so. Output: 0.96 0.93
positive
Input: Meyjes’s movie
to take all the output pairs and make them into the
same language.
Translation en-fr Input: account Output: compte 0.85 0.88
Input: rice Output: riz
Input: hardware Output: arme à feu

31
Published as a conference paper at ICLR 2023

Table 14: Test accuracies of best OpenAI Codex instructions with APE under six selected tasks

Task Instruction Prompt-only In-context

Antonyms write the opposite of the input. 0.83 0.84
read the two sentences and determine which one is
Cause Selection the cause and which one is the effect. If the first 0.76 0.96
sentence is the cause, write the first sentence.
write the output for each input by reversing the
Passivization order of the words in the input and changing the 1.00 1.00
verb to the passive voice.
Second Letter write the second letter of the input. 0.77 0.73
write a program that takes a movie review as in-
put and outputs a positive or negative sentiment.
Sentiment 0.91 0.95
The program should be able to distinguish between
positive and negative reviews.
write the French word for the English word. If
Translation en-fr you don’t know the French word, write the English 0.81 0.87
word.

Table 15: Test accuracies of best GLM-130B instructions with APE under six selected tasks

Task Instruction Prompt-only In-context

Antonyms generate the opposites. 0.82 0.83
Cause Selection read each sentence aloud. 0.48 0.80
Passivization read the input sentence. 0.64 1.00
Second Letter find the letter on each of its inputs. 0.22 0.39
Sentiment give them either positive or negative. 0.88 0.92
Translation en-fr translate English words into French. 0.75 0.87

32
Published as a conference paper at ICLR 2023

Table 16: Test accuracies of best APE GPT-3 instructions to prompt itself under six selected tasks

Task Instruction Prompt-only In-context

to translate the input word into its own antonym.
Thus, the correct answer to each input was the
Antonyms opposite word in the input word’s "opposite pair." 0.79 0.81
Inputs and outputs both had opposite pairs (except
for the first one
"Write a short story with the given inputs."
Inputs: Sentence 1: The door was locked. Sen-
Cause Selection tence 2: The man climbed in through the window. 0.36 0.76
Output: The door was locked. The man climbed in
through
input: The authors avoided the banker. Output:
The banker was avoided by the authors.
Passivization The instruction was: Input: The scientists encour- 1.00 1.00
aged the artists. Input: The artists were encouraged
by the scientists. Input
to find a word that rhymes with every input, and I
found out that the word "foible" rhymes with every
input word.
Second Letter 0.42 0.42
Input: defiance Output: a
Input: horse Output: e
Input
"describe your reaction to the movie "Julie & Ju-
lia", in one to five sentences." Output: positive
Sentiment 0.91 0.94
Input: Total crap. Output: negative
Input: Uplifting and funny. Output: positive
âœThink of the output as the subject of the verb in
the sentence.â Outputs and inputs were in French,
Translation en-fr 0.85 0.83
I gave the English translations. Here is my take:
Input: process Output: procès

33
Published as a conference paper at ICLR 2023

F A DDITIONAL V ISUALIZATIONS

Visualization Hyperparameters As we tuned the hyperparameters of APE including the number

of proposals generated per demonstration and the number of demonstrations per random seed, we
discovered better ones for instruction induction. We re-evaluated APE on 5 tasks, giving human-level
performance on all 24 of 24 instruction induction tasks. The additional visualizations below were
based on a previous iteration of APE which only reached human level on 19 of 24 tasks. The mean
test accuracy differences for those 5 tasks are summarized in Table 17.

Table 17: APE hyperparameter tuning improvements on instruction induction.

Task Name APE (Old) Accuracy, Mean APE (New) Accuracy, Mean APE (New) - Human
Second Letter 0.596 0.8 0.034
Pluralization 0.984 0.996 -0.004
Passivization 0.622 1 0.001
Sentence Similarity 0.186 0.256 -0.01
Membership 0.126 0.612 -0.001

Instruction (zero-shot) Only In-context Only Instruction (zero-shot) + In-context Instruction (few-shot) + In-context
1

0
Antonyms Cause Selection Common Concept Diff First Letter Formality Large Animal List Letters
Execution Accuracy

0
Membership Negation Number to Word Passivization Pluralization Rhymes Second Letter Sentence Similarity
1

0
Sentiment Starting With Sum Synonyms Translation en-de Translation en-es Translation en-fr Word in Context

Figure 14: Few-shot in-context test accuracy of best performing instructions selected using few-shot
execution accuracy on 24 Instruction Induction tasks.

InstructGPT CODEX OPT GLM

1
Execution Accuracy

0
Antonyms Cause Selection Passivization Second Letter Sentiment Translation en-fr

Figure 15: Zero-shot test accuracy on 6 Instruction Induction tasks. We compare the different models’
ability to propose instructions and use the InstructGPT for selection and execution.

34
Published as a conference paper at ICLR 2023

InstructGPT CODEX OPT GLM

1

Execution Accuracy

0
Antonyms Cause Selection Passivization Second Letter Sentiment Translation en-fr

Figure 16: Few-shot test accuracy on 6 Instruction Induction tasks. We compare the different models’
ability to propose instructions and use the InstructGPT for selection and execution.

35
Published as a conference paper at ICLR 2023

Figure 17: Zero-shot test accuracy on 6 Instruction Induction tasks. We investigate the transfer ability
of the APE instruction to a different model not involved during instruction generation and selection.

Figure 18: Zero-shot test accuracy of best performing instructions on 6 Instruction Induction tasks.
We investigate the transfer ability of the APE instruction to a different model not involved during
instruction generation and selection.

36
Published as a conference paper at ICLR 2023

Figure 19: Few-shot test accuracy on 6 Instruction Induction tasks. We investigate the transfer ability
of the APE instruction to a different model not involved during instruction generation and selection.

Figure 20: Few-shot test accuracy of best performing instructions on 6 Instruction Induction tasks.
We investigate the transfer ability of the APE instruction to a different model not involved during
instruction generation and selection.

37
Published as a conference paper at ICLR 2023

GPT-3_S GPT-3_M GPT-3_L GPT-3_XL InstructGPT_S InstructGPT_M InstructGPT_L InstructGPT_XL

1

0
First Letter Second Letter List Letters Starting With Pluralization Passivization Sentiment Sentence Similarity
Execution Accuracy

0
Word in Context Negation Antonyms Synonyms Membership Rhymes Large Animal Cause Selection
1

0
Common Concept Formality Sum Diff Number to Word Translation en-de Translation en-es Translation en-fr

Figure 23: Zero-shot test accuracy on 24 Instruction Induction tasks using eight different LLMs.

Forward (Template 1) Insert (Template 1) Insert (Template 2)

1
Execution Accuracy

0
Passivization Second Letter Starting With Sentence Similarity Synonyms Membership

Figure 21: Zero-shot test accuracy on 6 Instruction Induction tasks. We compare the performance
of different templates used to propose instruction. Insert Template 1 is adapted from instruction
induction, while Insert Template 2 is from TruthfulQA.

Forward (Template 1) Insert (Template 1) Insert (Template 2)

1
Execution Accuracy

0
Passivization Second Letter Starting With Sentence Similarity Synonyms Membership

Figure 22: Few-shot test accuracy on 6 Instruction Induction tasks. We compare the performance of
different templates used to propose instruction. Insert Template 1 is adpted from instruction induction,
while Insert Template 2 is from TruthfulQA.

38
Published as a conference paper at ICLR 2023

logp_forward logp_insert exec_forward exec_insert

1

0
Antonyms Cause Selection Common Concept Diff First Letter Formality Large Animal List Letters
Execution Accuracy

0
Membership Negation Number to Word Passivization Pluralization Rhymes Second Letter Sentence Similarity
1

0
Sentiment Starting With Sum Synonyms Translation en-de Translation en-es Translation en-fr Word in Context

Figure 24: Zero-shot test accuracy on 24 Instruction Induction tasks using two different metrics and
two different LLM models.

logp_forward logp_insert exec_forward exec_insert

1

0
Antonyms Cause Selection Common Concept Diff First Letter Formality Large Animal List Letters
Execution Accuracy

0
Membership Negation Number to Word Passivization Pluralization Rhymes Second Letter Sentence Similarity
1

0
Sentiment Starting With Sum Synonyms Translation en-de Translation en-es Translation en-fr Word in Context

Figure 25: In-Context learning without instruction on 24 Instruction Induction tasks using two
different metrics and two different LLM models.

logp_forward logp_insert exec_forward exec_insert

1

0
Antonyms Cause Selection Common Concept Diff First Letter Formality Large Animal List Letters
Execution Accuracy

0
Membership Negation Number to Word Passivization Pluralization Rhymes Second Letter Sentence Similarity
1

0
Sentiment Starting With Sum Synonyms Translation en-de Translation en-es Translation en-fr Word in Context

Figure 26: Test accuracy of in-Context learning with instruction on 24 Instruction Induction tasks
using two different metrics and two different LLM models.

39
Published as a conference paper at ICLR 2023

GPT-3 (350M) GPT-3 (6.7B) InstructGPT (350M) InstructGPT (6.7B)

GPT-3 (1.3B) GPT-3 (175B) InstructGPT (1.3B) InstructGPT (175B)
1.0
% instructions with accuracy >

2
0.8 10

0.6

Count
1
0.4 10

0.2
0
10
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Test accuracy ( ) Test accuracy ( )
Figure 27: Survival function and the histogram of test accuracy on a simple task (i.e. Pluralization)

GPT-3 (350M) GPT-3 (6.7B) InstructGPT (350M) InstructGPT (6.7B)

GPT-3 (1.3B) GPT-3 (175B) InstructGPT (1.3B) InstructGPT (175B)
1.0
% instructions with accuracy >

2
0.8 10

0.6
Count

1
0.4 10

0.2
0
10
0.0
0.0 0.2 0.4 0.6 0.0 0.2 0.4 0.6
Test accuracy ( ) Test accuracy ( )
Figure 28: Survival function and the histogram of test accuracy on a challenging task (i.e. Start With)

40
Published as a conference paper at ICLR 2023

Start 1 2 3 4 5
1.0
2
10
% instructions with accuracy >

0.8

0.6

Count
1
10
0.4

0.2
0
10
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Train accuracy ( ) Train accuracy ( )
Figure 29: Iterative Monte Carlo search improves the quality of the instruction candidates at each
round. Task: Antonyms.

Start 1 2 3 4 5
1.0
% instructions with accuracy >

0.8

0.6 1
Count

10

0.4

0.2
0
10
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Train accuracy ( ) Train accuracy ( )
Figure 30: Iterative Monte Carlo search improves the quality of the instruction candidates at each
round. Task: Cause Selection.

41
Published as a conference paper at ICLR 2023

Start 1 2 3 4 5
1.0
2
% instructions with accuracy >

10
0.8

0.6

Count
1
10
0.4

0.2
0
10
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Train accuracy ( ) Train accuracy ( )
Figure 31: Iterative Monte Carlo search improves the quality of the instruction candidates at each
round. Task: Passivization.

Start 1 2 3 4 5
1.0 10
2
% instructions with accuracy >

0.8

0.6
Count

1
10
0.4

0.2
0
10
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Train accuracy ( ) Train accuracy ( )
Figure 32: Iterative Monte Carlo search improves the quality of the instruction candidates at each
round. Task: Second Letter.

42
Published as a conference paper at ICLR 2023

Start 1 2 3 4 5
1.0
% instructions with accuracy >

2
10
0.8

0.6

Count
1
10
0.4

0.2
0
10
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Train accuracy ( ) Train accuracy ( )
Figure 33: Iterative Monte Carlo search improves the quality of the instruction candidates at each
round. Task: Sentiment.

Start 1 2 3 4 5
1.0 10
2
% instructions with accuracy >

0.8

0.6
Count

1
10
0.4

0.2
0
10
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Train accuracy ( ) Train accuracy ( )
Figure 34: Iterative Monte Carlo search improves the quality of the instruction candidates at each
round. Task: Translation en-fr.

43