tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
benchmark in Proceedings of the 28th International Conference on
Computational Linguistics pp 47624772 2020miniCodeProps a Minimal Benchmark for Proving
Code Properties
Evan Lohn
Carnegie Mellon University
evanlohncmueduSean Welleck
Carnegie Mellon University
welleckscmuedu
Abstract
Neural networks have shown initial promise in automating mathematical theorem
proving in proof assistants such as Lean The same proof assistants can be used to
verify the correctness of code by pairing code with specifications and proofs that
the specifications hold Automating the writing of code specifications and proofs
could lower the cost of verification or ambitiously enable a machine learning sys
tem to output provably correct code However it remains unclear whether current
neural theorem provers can automatically verify even relatively simple programs
We present miniCodeProps a benchmark of 177 program specifications in the
Lean proof assistant aimed at the subproblem of automatically generating a proof
for a provided program and specification miniCodeProps contains specifications
about simple selfcontained programs eg lists natural numbers binary trees
with varied proof difficulty Despite its simplicity miniCodeProps is challenging
for current LLMbased provers which succeed in proving about 25 percent of the
specifications We publicly release miniCodeProps as a benchmark for furthering
automated theorem proving in the context of formally verified code
1 Introduction
Writing code that meets a specification is a desirable yet difficult task In safetycritical contexts
users require stronger safeguards than code reviews and test cases provide To address this need the
formal methods community has produced a variety of tools for specifying and proving properties
of code 12345 One approach uses interactive theorem provers ITPs such as Isabelle 6
Coq 7 and Lean 8 In an ITP a user writes code a specification and a proof that the specification
holds Figure 1 The ITP checks each step of the proof as it is written and a complete proof means
that the specification is guaranteed to hold A complete proof thus gives guarantees on the underlying
code which is valuable in practice For instance the Lean proof assistant was recently used to verify
properties of Amazons Cedar Policy language1More speculatively future code generation models
could in principle output provably correct code if they are able to write code specifications and
proofs 9
Despite its promise interactive theorem proving remains a difficult task and better automationbe
it writing specifications proofs or combinations thereofcould help make it easier to verify code
Recently machine learning techniques based on large language models LLMs have shown promise
in automatically generating code 10 proving mathematical theorems in ITPs 11 and generating
proofs of properties from verification projects 12 However evaluating the capabilities of LLMs for
program verification remains a challenge For example benchmarks often evaluate across projects
with complex dependencies such as the CompCert compiler or Archive of Formal Proofs 1312
making it difficult to isolate and reason about a models core capabilities and weaknesses
1httpsgithubcomcedarpolicycedarspec
Preprint Under reviewarXiv240611915v1 csSE 16 Jun 2024Figure 1 Formal verification of code in an interactive proof assistant here Lean consists of writing
i code eg an implementation of a Tree ii a property that you want to verify eg that the size of
a balanced tree is odd and iii a proof that the code satisfies the property The underlying language
verifies that the proof is correct providing certainty that the property holds miniCodeProps focuses
on the important subproblem of proof generation given a property and associated code a model must
generate a proof eg shown here in green that the property holds miniCodeProps contains 177
properties about simple and selfcontained code blocks and measures a range of proving abilities
To address this need we introduce miniCodeProps a benchmark for evaluating the ability to prove
properties of relatively simple selfcontained programs in the Lean proof assistant miniCodeProps
contains 177 specifications about simple selfcontained programs involving eg lists natural
numbers binary trees with varied proof difficulty sourced by translating Haskell programs from
Tons of Inductive Programs 14 into Lean 4 Despite its simplicity miniCodeProps is challenging
for current LLMbased provers For example our best baseline approach based on GPT4 proved very
few specifications requiring proofs longer than a few lines We publicly release miniCodeProps as
a benchmark for furthering automated theorem proving in the context of formally verified code2
2 Related Work
Automating mathematical theorem proving Recently there has been wide interest in automating
mathematical theorem proving in interactive proof assistants see 1511 for surveys A typical
approach 16 is to train on a large corpus of mathematical proofs such as Leans Mathlib 1718
192021 A model learns to call automation that is designed for mathematics such as nonlinear
arithmetic tactics and to use definitions and theorems from within the corpus It is unclear whether
such methods transfer to the distribution of proofs encountered in program verification which may
rely on calling different automation using programspecific definitions or lemmas or different
proof strategies By focusing on Lean miniCodeProps allows for testing this transferability A
second class of methods design prompting strategies with mathematical problems in mind such as
conditioning generation on an informal proof 222324 or examples from Mathlib 25 We hope
thatminiCodeProps motivates development of similar methods for program verification
Automating formal code verification There is a rich history of developing automation for formal
verification of code see 4 for a survey For machine learning in interactive theorem proving
many methods that preceded large language models focused on Coq such as ProverBot9000 26
ASTactic 13 TacTok 27 Diva 28 and Passport 29 Recent work with LLMs includes exploring
prompting strategies in Coq 30 and Baldur 12 which explores proof generation and repair in
Isabelles Archive of Formal Proofs We are not aware of machine learningbased proof automation
targeting program verification in Lean despite the activity in automated mathematical proving in
2Our benchmark and baselines can be found here httpshuggingfacecodatasetselohnminiCodeProps
2Figure 2 miniCodeProps is sourced from Tons of Inductive Problems 14 a collection of programs
and specifications written in Haskell We translate the programs into Lean and write and prove
termination lemmas that are needed to state and prove properties of recursive functions
Lean One of our motivations is bridging this gap Finally some methods such as COPRA 24 aim to
be languageagnostic miniCodeProps could participate in evaluating and developing such systems
A second paradigm of verification is based on automated reasoning SATSMTbased languages such
as Dafny 1 and Verus 5 Rust Recent work explores LLMbased automation in these languages
such as Clover 9 for Dafny and 31 for Verus These are complementary to studying automated
ITP verification Finally Tons of Inductive Problems 14from which we derive our datatargets
automated reasoningbased verification while miniCodeProps targets interactive theorem proving
Interactive theorem proving benchmarks The CoqGym 13 benchmark tests on over 100 repos
itories in Coq while some papers use code properties from a single large complex repository such as
the CompCert compiler verification project in Coq 3224 which arguably tests different aspects of
automated code verification than those tested by miniCodeProps In Isabelle the Archive of Formal
Proofs contains some verificationrelated sections that are used for evaluation eg in Baldur 12
Benchmarks for automated theorem proving in Lean are focused on theorems from mathematical
domains For example miniF2F 33 contains 488 selfcontained easytostate theorems from math
competitions while ProofNet 34 contains selfcontained textbook problems miniF2Fs simplicity
and impact as a benchmark motivated the creation and naming of miniCodeProps
3 Benchmark Contents

benchmark in Proceedings of the 28th International Conference on

Computational Linguistics pp 47624772 2020miniCodeProps a Minimal Benchmark for Proving

Code Properties

Evan Lohn

Carnegie Mellon University

evanlohncmueduSean Welleck

Carnegie Mellon University

welleckscmuedu

Abstract

Neural networks have shown initial promise in automating mathematical theorem

proving in proof assistants such as Lean The same proof assistants can be used to

verify the correctness of code by pairing code with specifications and proofs that

the specifications hold Automating the writing of code specifications and proofs

could lower the cost of verification or ambitiously enable a machine learning sys

tem to output provably correct code However it remains unclear whether current

neural theorem provers can automatically verify even relatively simple programs

We present miniCodeProps a benchmark of 177 program specifications in the

Lean proof assistant aimed at the subproblem of automatically generating a proof

for a provided program and specification miniCodeProps contains specifications

about simple selfcontained programs eg lists natural numbers binary trees

with varied proof difficulty Despite its simplicity miniCodeProps is challenging

for current LLMbased provers which succeed in proving about 25 percent of the

specifications We publicly release miniCodeProps as a benchmark for furthering

automated theorem proving in the context of formally verified code

1 Introduction

Writing code that meets a specification is a desirable yet difficult task In safetycritical contexts

users require stronger safeguards than code reviews and test cases provide To address this need the

formal methods community has produced a variety of tools for specifying and proving properties

of code 12345 One approach uses interactive theorem provers ITPs such as Isabelle 6

Coq 7 and Lean 8 In an ITP a user writes code a specification and a proof that the specification

holds Figure 1 The ITP checks each step of the proof as it is written and a complete proof means

that the specification is guaranteed to hold A complete proof thus gives guarantees on the underlying

code which is valuable in practice For instance the Lean proof assistant was recently used to verify

properties of Amazons Cedar Policy language1More speculatively future code generation models

could in principle output provably correct code if they are able to write code specifications and

proofs 9

Despite its promise interactive theorem proving remains a difficult task and better automationbe

it writing specifications proofs or combinations thereofcould help make it easier to verify code

Recently machine learning techniques based on large language models LLMs have shown promise

in automatically generating code 10 proving mathematical theorems in ITPs 11 and generating

proofs of properties from verification projects 12 However evaluating the capabilities of LLMs for

program verification remains a challenge For example benchmarks often evaluate across projects

with complex dependencies such as the CompCert compiler or Archive of Formal Proofs 1312

making it difficult to isolate and reason about a models core capabilities and weaknesses

1httpsgithubcomcedarpolicycedarspec

Preprint Under reviewarXiv240611915v1 csSE 16 Jun 2024Figure 1 Formal verification of code in an interactive proof assistant here Lean consists of writing

i code eg an implementation of a Tree ii a property that you want to verify eg that the size of

a balanced tree is odd and iii a proof that the code satisfies the property The underlying language

verifies that the proof is correct providing certainty that the property holds miniCodeProps focuses

on the important subproblem of proof generation given a property and associated code a model must

generate a proof eg shown here in green that the property holds miniCodeProps contains 177

properties about simple and selfcontained code blocks and measures a range of proving abilities

To address this need we introduce miniCodeProps a benchmark for evaluating the ability to prove

properties of relatively simple selfcontained programs in the Lean proof assistant miniCodeProps

contains 177 specifications about simple selfcontained programs involving eg lists natural

numbers binary trees with varied proof difficulty sourced by translating Haskell programs from

Tons of Inductive Programs 14 into Lean 4 Despite its simplicity miniCodeProps is challenging

for current LLMbased provers For example our best baseline approach based on GPT4 proved very

few specifications requiring proofs longer than a few lines We publicly release miniCodeProps as

a benchmark for furthering automated theorem proving in the context of formally verified code2

2 Related Work

Automating mathematical theorem proving Recently there has been wide interest in automating

mathematical theorem proving in interactive proof assistants see 1511 for surveys A typical

approach 16 is to train on a large corpus of mathematical proofs such as Leans Mathlib 1718

192021 A model learns to call automation that is designed for mathematics such as nonlinear

arithmetic tactics and to use definitions and theorems from within the corpus It is unclear whether

such methods transfer to the distribution of proofs encountered in program verification which may

rely on calling different automation using programspecific definitions or lemmas or different

proof strategies By focusing on Lean miniCodeProps allows for testing this transferability A

second class of methods design prompting strategies with mathematical problems in mind such as

conditioning generation on an informal proof 222324 or examples from Mathlib 25 We hope

thatminiCodeProps motivates development of similar methods for program verification

Automating formal code verification There is a rich history of developing automation for formal

verification of code see 4 for a survey For machine learning in interactive theorem proving

many methods that preceded large language models focused on Coq such as ProverBot9000 26

ASTactic 13 TacTok 27 Diva 28 and Passport 29 Recent work with LLMs includes exploring

prompting strategies in Coq 30 and Baldur 12 which explores proof generation and repair in

Isabelles Archive of Formal Proofs We are not aware of machine learningbased proof automation

targeting program verification in Lean despite the activity in automated mathematical proving in

2Our benchmark and baselines can be found here httpshuggingfacecodatasetselohnminiCodeProps

2Figure 2 miniCodeProps is sourced from Tons of Inductive Problems 14 a collection of programs

and specifications written in Haskell We translate the programs into Lean and write and prove

termination lemmas that are needed to state and prove properties of recursive functions

Lean One of our motivations is bridging this gap Finally some methods such as COPRA 24 aim to

be languageagnostic miniCodeProps could participate in evaluating and developing such systems

A second paradigm of verification is based on automated reasoning SATSMTbased languages such

as Dafny 1 and Verus 5 Rust Recent work explores LLMbased automation in these languages

such as Clover 9 for Dafny and 31 for Verus These are complementary to studying automated

ITP verification Finally Tons of Inductive Problems 14from which we derive our datatargets

automated reasoningbased verification while miniCodeProps targets interactive theorem proving

Interactive theorem proving benchmarks The CoqGym 13 benchmark tests on over 100 repos

itories in Coq while some papers use code properties from a single large complex repository such as

the CompCert compiler verification project in Coq 3224 which arguably tests different aspects of

automated code verification than those tested by miniCodeProps In Isabelle the Archive of Formal

Proofs contains some verificationrelated sections that are used for evaluation eg in Baldur 12

Benchmarks for automated theorem proving in Lean are focused on theorems from mathematical

domains For example miniF2F 33 contains 488 selfcontained easytostate theorems from math

competitions while ProofNet 34 contains selfcontained textbook problems miniF2Fs simplicity

and impact as a benchmark motivated the creation and naming of miniCodeProps

3 Benchmark Contents