uiuc-kang-lab/agentic-benchmark-assessments

benchmark string	check string	score int64	reason string
SWE-Bench-Lancer	I.d.1	1	As discussed in Section 1 of the paper, the benchmark uses a set of test cases that are verified for correctness and quality by human experts.
SWE-Bench-Lancer	I.d.2	0	The benchmark does not use objective metrics to measure the quality of test cases.
SWE-Bench-Lancer	I.f.2	1	As discussed in Section 1, the end-to-end testing is designed to simulate the entire user workflow.
SWE-Bench-Lancer	I.f.3	0	The test cases use hard-coded timeouts, which may lead to non-deterministic results if the system is slow or unresponsive.
SWE-Bench-Lancer	II.1	1	The package dependencies are specified in the repository of each task.
SWE-Bench-Lancer	II.2	1	The benchmark does not require any external APIs.
SWE-Bench-Lancer	II.3	1	The benchmark does not require any external APIs.
SWE-Bench-Lancer	II.4	1	The benchmark uses docker containers to isolate the environment, and the state is cleared between runs.
SWE-Bench-Lancer	II.5	0	The agent can access the file system where the test cases are stored, which may lead to the agent accessing the ground truth information.
SWE-Bench-Lancer	II.6	1	The environment setup is static and does not change over time.
SWE-Bench-Lancer	II.7	1	The ground-truth test cases are taken from GitHub repositories, which are verified by expert developers.
SWE-Bench-Lancer	II.8	1	Each task represents a real-world software issue with a corresponding patch, which are solvable by the agent.
SWE-Bench-Lancer	II.9	1	The benchmark uses existing patches as ground truth, which can be considered as an Oracle solver.
SWE-Bench-Lancer	II.10	0	The benchmark does not handle the isolation between the agent and test cases properly. The test cases are stored not only in a file system that the agent can access, but also in a ZIP file that agent can read the directory structure and update files.
SWE-Bench-Lancer	III.1	1	The benchmark is open-sourced and available on GitHub.
SWE-Bench-Lancer	III.2	1	The benchmark provides an open-source evaluation harness for users.
SWE-Bench-Lancer	III.3	1	The benchmark maintains a private test set.
SWE-Bench-Lancer	III.4	0	The report does not discuss any measures or plans for consistent update.
SWE-Bench-Lancer	III.5	1	Such a relationship is clearly stated in Section 2 of the paper.
SWE-Bench-Lancer	III.6	1	As shown in Section 3, the benchmark is designed to evaluate the LLM model.
SWE-Bench-Lancer	III.7	1	The benchmark uses end-to-end testing to mitigate grader hacking.
SWE-Bench-Lancer	III.8	1	The benchmark discusses the potential impact of grader hacking in Section 1 and Appendix A.7.
SWE-Bench-Lancer	III.9	0	The benchmark does not include any quantitative analysis to assess the impact of grader hacking.
SWE-Bench-Lancer	III.10	0	The benchmark does not report any metrics about statistical significance.
SWE-Bench-Lancer	III.11	0	The benchmark does not provide any guidance on interpreting results with eval flaws.
SWE-Bench-Lancer	III.12	0	The benchmark does not report results of non-AI baselines.
SWE-Bench-Lancer	III.13	0	The benchmark does not report results of trivial agents.
Bird-Bench	I.d.1	1	As discussed in Section 3.4 of the paper, the validity of the database is verified by executing the ground-truth query.
Bird-Bench	I.d.2	0	The paper does not use objective metrics to measure the usefulness and completeness of the database or ground-truth queries.
Bird-Bench	I.f.2	0	The paper does not provide any information about the coverage of the database or ground-truth queries.
Bird-Bench	I.f.3	1	Executing SQL queries on a database is deterministic, and the paper does not mention any non-deterministic behavior.
Bird-Bench	II.1	1	The task instruction in Figure 9 speficies the SQL language is SQLite.
Bird-Bench	II.2	1	No external API is required for the evaluation of the benchmark.
Bird-Bench	II.3	1	No external API is required for the evaluation of the benchmark.
Bird-Bench	II.4	0	Databse file is neither opened in a read-only mode nor re-initialized between runs. This may lead to unexpected data manipulation by the agent.
Bird-Bench	II.5	1	Agent cannot access the host file system.
Bird-Bench	II.6	1	The environment setup is static and does not change over time.
Bird-Bench	II.7	0	As discussed in Section 3.4 of the paper, the correctness of the query is not fully verified, especially for the SQL queries that two annotators reach a concensus on.
Bird-Bench	II.8	0	The ambiguity of the SQL queries is not fully verified.
Bird-Bench	II.9	0	The Benchmark does not include an Oracle solver that can automatically solve all text-to-SQL tasks.
Bird-Bench	II.10	1	No vulnerabilities are found in the implementation of the benchmark.
Bird-Bench	III.1	1	The benchmark is open-sourced and available on GitHub.
Bird-Bench	III.2	1	The benchmark provides an open-source evaluation harness for users.
Bird-Bench	III.3	1	The benchmark has a private test set.
Bird-Bench	III.4	0	The benchmark does not discuss any measures or plans for consistent update.
Bird-Bench	III.5	1	It is clearly discussed in Section 2 of the paper.
Bird-Bench	III.6	1	It is clearly discussed in Section 2 of the paper.
Bird-Bench	III.7	0	No efforts are made to mitigate errors when both annotators make the same mistake.
Bird-Bench	III.8	0	The paper does not discuss the potential impact of annotation errors.
Bird-Bench	III.9	0	The paper does not analyze the quantitative impact of annotation errors.
Bird-Bench	III.10	0	The paper does not report any metrics about statistical significance.
Bird-Bench	III.11	0	The paper does not provide any guidance on interpreting results with eval flaws.
Bird-Bench	III.12	1	The paper reports the results of human experts.
Bird-Bench	III.13	0	The paper does not report the results of any trivial agents.
Cy-Bench	I.h.1	1	The specific format required for the answer is provided in the task description.
Cy-Bench	I.h.2	1	The ground truth is complex enough to prevent trivial guessing.
Cy-Bench	II.1	1	Agents are granted access to using all tools. The versions of tools can be checked by the agent.
Cy-Bench	II.2	1	The benchmark does not require any external APIs.
Cy-Bench	II.3	1	The benchmark does not require any external APIs.
Cy-Bench	II.4	1	The benchmark uses docker containers to isolate the environment, and the state is cleared between runs.
Cy-Bench	II.5	1	The agent cannot directly access the container which contains the ground truth.
Cy-Bench	II.6	1	The environment setup is static and does not change over time.
Cy-Bench	II.7	1	As shown in Section 3.3 of the paper, the ground truth is verified by human.
Cy-Bench	II.8	1	As shown in Section 3.3 of the paper, each task is verified to be solvable.
Cy-Bench	II.9	1	As shown in Section 3.3 of the paper, the benchmark includes an Oracle solver that can automatically solve all tasks.
Cy-Bench	II.10	1	No vulnerabilities are found in the implementation of the benchmark.
Cy-Bench	III.1	1	The benchmark is open-sourced and available on GitHub.
Cy-Bench	III.2	1	The benchmark provides an open-source evaluation harness for users.
Cy-Bench	III.3	0	The benchmark does not contain measures to prevent data contamination.
Cy-Bench	III.4	0	The report does not discuss plans to consistently update tasks over time.
Cy-Bench	III.5	1	Such a relationship is clearly stated in Section 1 of the paper.
Cy-Bench	III.6	1	As shown in Section 1, the benchmark is designed to evaluate both agent frameworks and LLM models.
Cy-Bench	III.7	1	Annotation flaws are mitigated by developing verifiable tasks.
Cy-Bench	III.8	1	No unavoidable flaws are identified in the benchmark.
Cy-Bench	III.9	1	No unavoidable flaws are identified in the benchmark.
Cy-Bench	III.10	0	The report does not include any metrics about statistical significance.
Cy-Bench	III.11	1	No evaluation flaws are identified in the benchmark.
Cy-Bench	III.12	1	Human peerformance is reported in Section 5 of the paper.
Cy-Bench	III.13	0	The report does not report results of trivial agents.
SWE-Bench-Verified	I.d.1	1	Test cases are directly taken from GitHub repositories, and the paper does not mention any verification process.
SWE-Bench-Verified	I.d.2	0	The paper does not use objective metrics to measure quality of test cases.
SWE-Bench-Verified	II.1	1	The versions of package dependencies are specified in the repository.
SWE-Bench-Verified	II.2	1	The benchmark does not require any external APIs.
SWE-Bench-Verified	II.3	1	The benchmark does not require any external APIs.
SWE-Bench-Verified	II.4	1	The benchmark uses docker containers to isolate the environment, and the state is cleared between runs.
SWE-Bench-Verified	II.5	1	The agent cannot access the host file system, and the ground truth is not accessible to the agent.
SWE-Bench-Verified	II.6	1	The environment setup is static and does not change over time.
SWE-Bench-Verified	II.7	1	The ground-truth patches are taken from GitHub repositories, which is verified by expert developers.
SWE-Bench-Verified	II.8	1	Each task represents a real-world GitHub issue and a corresponding pull request, which are solvable by the agent.
SWE-Bench-Verified	II.9	1	Pull requests from GitHub are used as ground truth, which can be considered as an Oracle solver.
SWE-Bench-Verified	II.10	1	No vulnerabilities are found in the implementation of the benchmark, and the evaluation process is secure.
SWE-Bench-Verified	III.1	1	The benchmark is open-sourced and available on GitHub.
SWE-Bench-Verified	III.2	1	The benchmark provides an open-source evaluation harness for users.
SWE-Bench-Verified	III.3	0	The benchmark does not discuss measures to prevent data contamination.
SWE-Bench-Verified	III.4	0	The benchmark does not discuss plans to consistently update tasks over time.
SWE-Bench-Verified	III.5	1	Such a relationship is clearly stated in Section 2 of the paper.
SWE-Bench-Verified	III.6	1	The benchmark is designed to evaluate both the model and the agent framework, as discussed in Section 5 of the paper.
SWE-Bench-Verified	III.7	0	The benchmark does not discuss any efforts to prevent, identify, and correct flaws.
SWE-Bench-Verified	III.8	0	The benchmark does not discuss the potential impact of unavoidable flaws.
SWE-Bench-Verified	III.9	0	The benchmark does not include quantitative analysis to assess the impact of unavoidable flaws.

SWE-Bench-Lancer

I.d.1

1

As discussed in Section 1 of the paper, the benchmark uses a set of test cases that are verified for correctness and quality by human experts.

SWE-Bench-Lancer

I.d.2

0

The benchmark does not use objective metrics to measure the quality of test cases.

SWE-Bench-Lancer

I.f.2

1

As discussed in Section 1, the end-to-end testing is designed to simulate the entire user workflow.

SWE-Bench-Lancer

I.f.3

0

The test cases use hard-coded timeouts, which may lead to non-deterministic results if the system is slow or unresponsive.

SWE-Bench-Lancer

II.1

1

The package dependencies are specified in the repository of each task.

SWE-Bench-Lancer

II.2

1

The benchmark does not require any external APIs.

SWE-Bench-Lancer

II.3

1

The benchmark does not require any external APIs.

SWE-Bench-Lancer

II.4

1

The benchmark uses docker containers to isolate the environment, and the state is cleared between runs.

SWE-Bench-Lancer

II.5

0

The agent can access the file system where the test cases are stored, which may lead to the agent accessing the ground truth information.

SWE-Bench-Lancer

II.6

1

The environment setup is static and does not change over time.

SWE-Bench-Lancer

II.7

1

The ground-truth test cases are taken from GitHub repositories, which are verified by expert developers.

SWE-Bench-Lancer

II.8

1

Each task represents a real-world software issue with a corresponding patch, which are solvable by the agent.

SWE-Bench-Lancer

II.9

1

The benchmark uses existing patches as ground truth, which can be considered as an Oracle solver.

SWE-Bench-Lancer

II.10

0

The benchmark does not handle the isolation between the agent and test cases properly. The test cases are stored not only in a file system that the agent can access, but also in a ZIP file that agent can read the directory structure and update files.

SWE-Bench-Lancer

III.1

1

The benchmark is open-sourced and available on GitHub.

SWE-Bench-Lancer

III.2

1

The benchmark provides an open-source evaluation harness for users.

SWE-Bench-Lancer

III.3

1

The benchmark maintains a private test set.

SWE-Bench-Lancer

III.4

0

The report does not discuss any measures or plans for consistent update.

SWE-Bench-Lancer

III.5

1

Such a relationship is clearly stated in Section 2 of the paper.

SWE-Bench-Lancer

III.6

1

As shown in Section 3, the benchmark is designed to evaluate the LLM model.

SWE-Bench-Lancer

III.7

1

The benchmark uses end-to-end testing to mitigate grader hacking.

SWE-Bench-Lancer

III.8

1

The benchmark discusses the potential impact of grader hacking in Section 1 and Appendix A.7.

SWE-Bench-Lancer

III.9

0

The benchmark does not include any quantitative analysis to assess the impact of grader hacking.

SWE-Bench-Lancer

III.10

0

The benchmark does not report any metrics about statistical significance.

SWE-Bench-Lancer

III.11

0

The benchmark does not provide any guidance on interpreting results with eval flaws.

SWE-Bench-Lancer

III.12

0

The benchmark does not report results of non-AI baselines.

SWE-Bench-Lancer

III.13

0

The benchmark does not report results of trivial agents.

Bird-Bench

I.d.1

1

As discussed in Section 3.4 of the paper, the validity of the database is verified by executing the ground-truth query.

Bird-Bench

I.d.2

0

The paper does not use objective metrics to measure the usefulness and completeness of the database or ground-truth queries.

Bird-Bench

I.f.2

0

The paper does not provide any information about the coverage of the database or ground-truth queries.

Bird-Bench

I.f.3

1

Executing SQL queries on a database is deterministic, and the paper does not mention any non-deterministic behavior.

Bird-Bench

II.1

1

The task instruction in Figure 9 speficies the SQL language is SQLite.

Bird-Bench

II.2

1

No external API is required for the evaluation of the benchmark.

Bird-Bench

II.3

1

No external API is required for the evaluation of the benchmark.

Bird-Bench

II.4

0

Databse file is neither opened in a read-only mode nor re-initialized between runs. This may lead to unexpected data manipulation by the agent.

Bird-Bench

II.5

1

Agent cannot access the host file system.

Bird-Bench

II.6

1

The environment setup is static and does not change over time.

Bird-Bench

II.7

0

As discussed in Section 3.4 of the paper, the correctness of the query is not fully verified, especially for the SQL queries that two annotators reach a concensus on.

Bird-Bench

II.8

0

The ambiguity of the SQL queries is not fully verified.

Bird-Bench

II.9

0

The Benchmark does not include an Oracle solver that can automatically solve all text-to-SQL tasks.

Bird-Bench

II.10

1

No vulnerabilities are found in the implementation of the benchmark.

Bird-Bench

III.1

1

The benchmark is open-sourced and available on GitHub.

Bird-Bench

III.2

1

The benchmark provides an open-source evaluation harness for users.

Bird-Bench

III.3

1

The benchmark has a private test set.

Bird-Bench

III.4

0

The benchmark does not discuss any measures or plans for consistent update.

Bird-Bench

III.5

1

It is clearly discussed in Section 2 of the paper.

Bird-Bench

III.6

1

It is clearly discussed in Section 2 of the paper.

Bird-Bench

III.7

0

No efforts are made to mitigate errors when both annotators make the same mistake.

Bird-Bench

III.8

0

The paper does not discuss the potential impact of annotation errors.

Bird-Bench

III.9

0

The paper does not analyze the quantitative impact of annotation errors.

Bird-Bench

III.10

0

The paper does not report any metrics about statistical significance.

Bird-Bench

III.11

0

The paper does not provide any guidance on interpreting results with eval flaws.

Bird-Bench

III.12

1

The paper reports the results of human experts.

Bird-Bench

III.13

0

The paper does not report the results of any trivial agents.

Cy-Bench

I.h.1

1

The specific format required for the answer is provided in the task description.

Cy-Bench

I.h.2

1

The ground truth is complex enough to prevent trivial guessing.

Cy-Bench

II.1

1

Agents are granted access to using all tools. The versions of tools can be checked by the agent.

Cy-Bench

II.2

1

The benchmark does not require any external APIs.

Cy-Bench

II.3

1

The benchmark does not require any external APIs.

Cy-Bench

II.4

1

The benchmark uses docker containers to isolate the environment, and the state is cleared between runs.

Cy-Bench

II.5

1

The agent cannot directly access the container which contains the ground truth.

Cy-Bench

II.6

1

The environment setup is static and does not change over time.

Cy-Bench

II.7

1

As shown in Section 3.3 of the paper, the ground truth is verified by human.

Cy-Bench

II.8

1

As shown in Section 3.3 of the paper, each task is verified to be solvable.

Cy-Bench

II.9

1

As shown in Section 3.3 of the paper, the benchmark includes an Oracle solver that can automatically solve all tasks.

Cy-Bench

II.10

1

No vulnerabilities are found in the implementation of the benchmark.

Cy-Bench

III.1

1

The benchmark is open-sourced and available on GitHub.

Cy-Bench

III.2

1

The benchmark provides an open-source evaluation harness for users.

Cy-Bench

III.3

0

The benchmark does not contain measures to prevent data contamination.

Cy-Bench

III.4

0

The report does not discuss plans to consistently update tasks over time.

Cy-Bench

III.5

1

Such a relationship is clearly stated in Section 1 of the paper.

Cy-Bench

III.6

1

As shown in Section 1, the benchmark is designed to evaluate both agent frameworks and LLM models.

Cy-Bench

III.7

1

Annotation flaws are mitigated by developing verifiable tasks.

Cy-Bench

III.8

1

No unavoidable flaws are identified in the benchmark.

Cy-Bench

III.9

1

No unavoidable flaws are identified in the benchmark.

Cy-Bench

III.10

0

The report does not include any metrics about statistical significance.

Cy-Bench

III.11

1

No evaluation flaws are identified in the benchmark.

Cy-Bench

III.12

1

Human peerformance is reported in Section 5 of the paper.

Cy-Bench

III.13

0

The report does not report results of trivial agents.

SWE-Bench-Verified

I.d.1

1

Test cases are directly taken from GitHub repositories, and the paper does not mention any verification process.

SWE-Bench-Verified

I.d.2

0

The paper does not use objective metrics to measure quality of test cases.

SWE-Bench-Verified

II.1

1

The versions of package dependencies are specified in the repository.

SWE-Bench-Verified

II.2

1

The benchmark does not require any external APIs.

SWE-Bench-Verified

II.3

1

The benchmark does not require any external APIs.

SWE-Bench-Verified

II.4

1

The benchmark uses docker containers to isolate the environment, and the state is cleared between runs.

SWE-Bench-Verified

II.5

1

The agent cannot access the host file system, and the ground truth is not accessible to the agent.

SWE-Bench-Verified

II.6

1

The environment setup is static and does not change over time.

SWE-Bench-Verified

II.7

1

The ground-truth patches are taken from GitHub repositories, which is verified by expert developers.

SWE-Bench-Verified

II.8

1

Each task represents a real-world GitHub issue and a corresponding pull request, which are solvable by the agent.

SWE-Bench-Verified

II.9

1

Pull requests from GitHub are used as ground truth, which can be considered as an Oracle solver.

SWE-Bench-Verified

II.10

1

No vulnerabilities are found in the implementation of the benchmark, and the evaluation process is secure.

SWE-Bench-Verified

III.1

1

The benchmark is open-sourced and available on GitHub.

SWE-Bench-Verified

III.2

1

The benchmark provides an open-source evaluation harness for users.

SWE-Bench-Verified

III.3

0

The benchmark does not discuss measures to prevent data contamination.

SWE-Bench-Verified

III.4

0

The benchmark does not discuss plans to consistently update tasks over time.

SWE-Bench-Verified

III.5

1

Such a relationship is clearly stated in Section 2 of the paper.

SWE-Bench-Verified

III.6

1

The benchmark is designed to evaluate both the model and the agent framework, as discussed in Section 5 of the paper.

SWE-Bench-Verified

III.7

0

The benchmark does not discuss any efforts to prevent, identify, and correct flaws.

SWE-Bench-Verified

III.8

0

The benchmark does not discuss the potential impact of unavoidable flaws.

SWE-Bench-Verified

III.9

0

The benchmark does not include quantitative analysis to assess the impact of unavoidable flaws.