tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
35 Availability and Usage
We have published our benchmark as a public dataset on Huggingface including a link to our
example benchmarking code The TIP repo our benchmark is based on uses the BSD3Clause
license permitting reuse and modification with license redistribution we therefore use the same
Given the information in Table 2 there are many possible tasks relevant for testing an agents proof
generation capabilities We propose the following as interesting evaluation modes
1Proof State only Past work has trained nextstep tactic prediction models taking only the
proof state provided by the proof assistant as input 16192135 This input is justifiable
in contexts such as mathematical theorem proving where most proofs involve a shared
mathematical library because agents will be exposed to most of the definitions and lemmas
used in the goal during training We expect this input to be insufficient for many proofs
about programs as the programs referenced in the proof state are only described in the file
context However this mode will test an agents ability to apply reasonable proof techniques
when lacking knowledge of certain terms in the proof state
7Model Medley Easy Termination Med Sorting Hard
llmstep Pythia28B 35 4486 128 063
llmstep Llemma7B 25 4686 228 063
llmstep ntpcontext13B 36 3886 028 063
GPT4turbo 4486 128 063
Table 3 miniCodeProps results Number of specifications proven when applying nextstep tactic
generation with LLMStep and full proof generation with GPT4 to the problem of verifying program
specifications The Medley section contains mostly of specifications that can be proven in several
lines Proofs of the sorting algorithm properties and termination lemmas are expected to require at
least tens of lines and hours of programmer effort
2Proof State Dependencies In this mode agents have access to all newly written ie not
from Lean core or Mathlib code and lemmas required to define the given code property
This can be accomplished with the deps and proof_state fields of each benchmark entry
3Proof State Full File Context In this mode the full contents of the files prior to the
property definition are made available Similar to Proof State Dependencies mode this
input form gives an agent access to all definitions in the proof state However full file
context mode requires the agent to filter its input for useful context and additionally may
give the agent access to potentially useful lemmas defined in the file that were not necessary
to define the property Further this mode most closely matches what a human expert must
do when proving such properties
4Feedback Refinement In this mode error feedback from potentially partial proof sugges
tions is used as input to continued attempts at proving the property until a given number of
attempts is reached In practice Lean is intended to be used interactively Even for simple
proofs it may take human experts several iterations of error feedback from Lean to construct
a correct proof Enabling this feedback loop for theorem proving agents allows more fair
comparison of current neural theorem proving approaches with human capabilities
4 Baselines
We use two main approaches full proof generation via fewshot prompting a language model and
nextstep tactic generation In the fullproof context the model generates one or more potential proofs
that are verified by the proof checker in our case the Lean 4 kernel For nextstep tactic prediction
we follow the common practice of taking the proof state and optionally file context as input and
returning suggestions for the next tactic to use in the proof Each suggestion is then given as input to
the Lean 4 kernel and the resulting proof state is used to prompt the model for the next tactic
Experimental Setup NextStep Tactic Generation Our experiments with nextstep tactic genera
tion used the LLMStep 35 framework modified to return cumulative log probabilities in addition to
next step suggestions from each backend language model The Pythia28B experiments were run
with proof state only as input to match the models finetuning set The remaining models received
proof state dependencies as input These experiments were run on a single A100 GPU running
Ubuntu 2004 with 30 CPU cores rented from LambdaLabs
Experimental Setup Full Proof Generation We employed a fewshot prompting approach Our
prompt contained examples of code properties and proofs of properties not seen in the benchmark
As such the evaluation mode used was proof state dependencies although in practice we opted to
use only the dependencies and the theorem statement itself All experiments were performed using
the OpenAI chat completions API with GPT4turbo using pass8
Results Our baseline results are displayed in Table 3 As expected our baseline approaches were
unable to prove complex function properties such as those in Section 343 However failures on
many of the comparatively simple Medley properties support the claim that current approaches to
neural theorem for mathematics transfer imperfectly to the domain of ITP code verification However
small measures of success on the termination lemmas indicate that neural theorem proving approaches
have the potential to improve on miniCodeProps in the near future
85 Discussion
Baselines Interpretation The baselines we present are a reasonable representation of the state of
neural theorem proving on code properties at the moment Current systems can prove the simpler
lemmas in the Medley section especially those with closer ties to math and are only able to prove
the simpler types of lemmas one might use in a larger proof in practice Proofs of more complex
properties requiring understanding larger bodies of code the Sorting section remain out of reach for
the time being
Scope We chose to source miniCodeProps from the TIP dataset to provide some assurance that the
code and properties involved were of interest to the broader code verification community However
we selected a subset of the properties in TIP we judged to be relevant to code along with termination
lemmas which is a subset of all code properties Clearly performing well on this subset is only
a proxy for the ability to perform well on all realistic code properties However we designed
miniCodeProps to target a minimal set of code properties that we would expect a reasonable
automated system to be able to prove
Societal Context The broader goal of our work is to facilitate the development of automated likely
neural theorem proving agents that can prove properties of code in ITPs Verification engineers
already draw from a wide variety of automated tools in their work An automated version of ITP
still requires a human in the loop to generate or validate the statement of the property to be proven
similar to the type of work these engineers perform while using SMTbased verification Arguably
the main benefit to society of the ITP approach is that it induces our automated tools to use different
reasoning patterns than the ones commonly used in Automated Reasoning SATSMT approaches
which may in turn lead to different subsets of formal guarantees that can be obtained in practice
6 Conclusion
miniCodeProps is intended as a meaningful benchmark for evaluating techniques that automate
ITPbased code verification It contains a diverse array of functions and properties from Tons
of Inductive Problems allowing the possibility of incremental progress in verifying the types of
programs contained in TIP We show that simple baseline approaches from mathematical neural
theorem proving currently fall short on much of our benchmark and hope that miniCodeProps
spurs development of future ITP code verification agents
References
1Rustan Leino Dafny An automatic program verifier for functional correctness In 16th
International Conference LPAR16 Dakar Senegal pages 348370 Springer Berlin Heidelberg
April 2010
2Rustan Leino and Philipp Rümmer A polymorphic intermediate verification language Design
and logical encoding In Conference Tools and Algorithms for the Construction and Analysis
of Systems 16th International Conference TACAS 2010 Held as Part of the Joint European

35 Availability and Usage

We have published our benchmark as a public dataset on Huggingface including a link to our

example benchmarking code The TIP repo our benchmark is based on uses the BSD3Clause

license permitting reuse and modification with license redistribution we therefore use the same

Given the information in Table 2 there are many possible tasks relevant for testing an agents proof

generation capabilities We propose the following as interesting evaluation modes

1Proof State only Past work has trained nextstep tactic prediction models taking only the

proof state provided by the proof assistant as input 16192135 This input is justifiable

in contexts such as mathematical theorem proving where most proofs involve a shared

mathematical library because agents will be exposed to most of the definitions and lemmas

used in the goal during training We expect this input to be insufficient for many proofs

about programs as the programs referenced in the proof state are only described in the file

context However this mode will test an agents ability to apply reasonable proof techniques

when lacking knowledge of certain terms in the proof state

7Model Medley Easy Termination Med Sorting Hard

llmstep Pythia28B 35 4486 128 063

llmstep Llemma7B 25 4686 228 063

llmstep ntpcontext13B 36 3886 028 063

GPT4turbo 4486 128 063

Table 3 miniCodeProps results Number of specifications proven when applying nextstep tactic

generation with LLMStep and full proof generation with GPT4 to the problem of verifying program

specifications The Medley section contains mostly of specifications that can be proven in several

lines Proofs of the sorting algorithm properties and termination lemmas are expected to require at

least tens of lines and hours of programmer effort

2Proof State Dependencies In this mode agents have access to all newly written ie not

from Lean core or Mathlib code and lemmas required to define the given code property

This can be accomplished with the deps and proof_state fields of each benchmark entry

3Proof State Full File Context In this mode the full contents of the files prior to the

property definition are made available Similar to Proof State Dependencies mode this

input form gives an agent access to all definitions in the proof state However full file

context mode requires the agent to filter its input for useful context and additionally may

give the agent access to potentially useful lemmas defined in the file that were not necessary

to define the property Further this mode most closely matches what a human expert must

do when proving such properties

4Feedback Refinement In this mode error feedback from potentially partial proof sugges

tions is used as input to continued attempts at proving the property until a given number of

attempts is reached In practice Lean is intended to be used interactively Even for simple

proofs it may take human experts several iterations of error feedback from Lean to construct

a correct proof Enabling this feedback loop for theorem proving agents allows more fair

comparison of current neural theorem proving approaches with human capabilities

4 Baselines

We use two main approaches full proof generation via fewshot prompting a language model and

nextstep tactic generation In the fullproof context the model generates one or more potential proofs

that are verified by the proof checker in our case the Lean 4 kernel For nextstep tactic prediction

we follow the common practice of taking the proof state and optionally file context as input and

returning suggestions for the next tactic to use in the proof Each suggestion is then given as input to

the Lean 4 kernel and the resulting proof state is used to prompt the model for the next tactic

Experimental Setup NextStep Tactic Generation Our experiments with nextstep tactic genera

tion used the LLMStep 35 framework modified to return cumulative log probabilities in addition to

next step suggestions from each backend language model The Pythia28B experiments were run

with proof state only as input to match the models finetuning set The remaining models received

proof state dependencies as input These experiments were run on a single A100 GPU running

Ubuntu 2004 with 30 CPU cores rented from LambdaLabs

Experimental Setup Full Proof Generation We employed a fewshot prompting approach Our

prompt contained examples of code properties and proofs of properties not seen in the benchmark

As such the evaluation mode used was proof state dependencies although in practice we opted to

use only the dependencies and the theorem statement itself All experiments were performed using

the OpenAI chat completions API with GPT4turbo using pass8

Results Our baseline results are displayed in Table 3 As expected our baseline approaches were

unable to prove complex function properties such as those in Section 343 However failures on

many of the comparatively simple Medley properties support the claim that current approaches to

neural theorem for mathematics transfer imperfectly to the domain of ITP code verification However

small measures of success on the termination lemmas indicate that neural theorem proving approaches

have the potential to improve on miniCodeProps in the near future

85 Discussion

Baselines Interpretation The baselines we present are a reasonable representation of the state of

neural theorem proving on code properties at the moment Current systems can prove the simpler

lemmas in the Medley section especially those with closer ties to math and are only able to prove

the simpler types of lemmas one might use in a larger proof in practice Proofs of more complex

properties requiring understanding larger bodies of code the Sorting section remain out of reach for

the time being

Scope We chose to source miniCodeProps from the TIP dataset to provide some assurance that the

code and properties involved were of interest to the broader code verification community However

we selected a subset of the properties in TIP we judged to be relevant to code along with termination

lemmas which is a subset of all code properties Clearly performing well on this subset is only

a proxy for the ability to perform well on all realistic code properties However we designed

miniCodeProps to target a minimal set of code properties that we would expect a reasonable

automated system to be able to prove

Societal Context The broader goal of our work is to facilitate the development of automated likely

neural theorem proving agents that can prove properties of code in ITPs Verification engineers

already draw from a wide variety of automated tools in their work An automated version of ITP

still requires a human in the loop to generate or validate the statement of the property to be proven

similar to the type of work these engineers perform while using SMTbased verification Arguably

the main benefit to society of the ITP approach is that it induces our automated tools to use different

reasoning patterns than the ones commonly used in Automated Reasoning SATSMT approaches

which may in turn lead to different subsets of formal guarantees that can be obtained in practice

6 Conclusion

miniCodeProps is intended as a meaningful benchmark for evaluating techniques that automate

ITPbased code verification It contains a diverse array of functions and properties from Tons

of Inductive Problems allowing the possibility of incremental progress in verifying the types of

programs contained in TIP We show that simple baseline approaches from mathematical neural

theorem proving currently fall short on much of our benchmark and hope that miniCodeProps

spurs development of future ITP code verification agents

References

1Rustan Leino Dafny An automatic program verifier for functional correctness In 16th

International Conference LPAR16 Dakar Senegal pages 348370 Springer Berlin Heidelberg

April 2010

2Rustan Leino and Philipp Rümmer A polymorphic intermediate verification language Design

and logical encoding In Conference Tools and Algorithms for the Construction and Analysis

of Systems 16th International Conference TACAS 2010 Held as Part of the Joint European