benchmark
string
check
string
score
int64
reason
string
GAIA
II.7
1
The data annotation process contains a verification step, as discussed in Section 3.4 of the paper.
GAIA
II.8
1
The data annotation process contains a verification step, as discussed in Section 3.4 of the paper.
GAIA
II.9
0
The benchmark does not include an Oracle solver that can automatically solve all tasks.
GAIA
II.10
1
No vulnerabilities are found in the implementation of the benchmark.
GAIA
III.1
1
The benchmark is open-sourced and available on HuggingFace.
GAIA
III.2
0
The benchmark does not provide an open-source evaluation harness for users.
GAIA
III.3
0
The benchmark does not contain measures to prevent data contamination.
GAIA
III.4
0
The report does not discuss plans to consistently update tasks over time.
GAIA
III.5
1
Such a relationship is clearly stated in Section 3 of the paper.
GAIA
III.6
1
As discussed in Section 3 of the paper, the benchmark is designed to evaluate LLM models.
GAIA
III.7
1
Section 5 of the paper discusses the efforts, including comparing evaluation with or without human in the loop.
GAIA
III.8
1
Section 6 discusses the potential impact of unavoidable flaws, such as a wrong reasoning trace resulting in a correct answer.
GAIA
III.9
0
The report does not include quantitative analysis to assess the impact of unavoidable flaws.
GAIA
III.10
0
The report does not include any metrics about statistical significance.
GAIA
III.11
0
The report does not provide any guidance on interpreting results with eval flaws.
GAIA
III.12
1
Human performance is reported in Section 4 of the paper.
GAIA
III.13
1
The report includes results of search engine, which can be considered a trivial agent.
OSWorld
I.g.1
1
As discussed in Section 3.2 of the paper, the ground truth is verified to include all states that can be achieved after a successful task completion.
OSWorld
I.g.2
0
The state check only verifies the relevant states for the tasks. Agents can potentially perform extra harmful actions that are not checked by the ground truth.
OSWorld
I.g.3
1
AS demonstrated in Section 3.2 of the paper, the ground truth involves complex state changes to a software or website.
OSWorld
II.1
1
No external tools are used in the benchmark. Versions of the environment are clearly specified in the README file of the repository.
OSWorld
II.2
1
No external APIs are used in the benchmark.
OSWorld
II.3
1
No external APIs are used in the benchmark.
OSWorld
II.4
1
The benchmark uses virtual machines to run the tasks, which ensures that all residual data or state are cleared between runs.
OSWorld
II.5
1
Agents and ground truth are isolated from each other via virtual machines.
OSWorld
II.6
0
The benchmark checks for HTML selectors (like class names or page titles) on live web pages.
OSWorld
II.7
1
As discussed in Section 3.2 of the paper, the ground truth is verified for correctness by human experts.
OSWorld
II.8
1
As discussed in Section 3.2 of the paper, each task is verified to be solvable by human experts.
OSWorld
II.9
0
The benchmark does not include an Oracle solver that can automatically solve all tasks.
OSWorld
II.10
1
No vulnerabilities are present in the implementation of the benchmark.
OSWorld
III.1
1
The benchmark is fully open-sourced, as the code is available on GitHub.
OSWorld
III.2
1
The benchmark offers an open-source evaluation harness for users.
OSWorld
III.3
0
The benchmark does not include measures to prevent data contamination.
OSWorld
III.4
0
The report does not include measures or plans to consistently update tasks over time.
OSWorld
III.5
1
Such a relationship is clearly stated in Section 2 of the paper.
OSWorld
III.6
1
As discussed in Section 2 of the paper, the evaluation subject is agent frameworks.
OSWorld
III.7
1
As discussed in Section 3.2 of the paper, the benchmark uses additional manual verification steps to prevent, identify, and correct flaws.
OSWorld
III.8
0
Safety issues of agents are discussed in Section 7 of the paper.
OSWorld
III.9
0
No quantitative analysis to assess the impact of unavoidable flaws is included in the report.
OSWorld
III.10
0
The report does not include metrics about statistical significance.
OSWorld
III.11
0
The report does not provide guidance on interpreting results with eval flaws.
OSWorld
III.12
1
Human performace is reported in Section 3.4 of the paper.
OSWorld
III.13
0
The report does not include results of trivial agents.
KernelBench
I.e.1
0
The fuzzer does not address potential edge cases, such as empty inputs.
KernelBench
I.e.2
0
Although the data type is specified, the fuzzer does not test different memory layouts, such as tensors with incontiguous memory layouts.
KernelBench
I.e.3
0
The fuzzer uses uniform sampling to generate inputs, which may not be sensitive to the code under testing. For example, the fuzzer may not generate positive inputs that trigger the `relu` function in the `torch` library.
KernelBench
II.1
0
The CUDA version is not specified in the default prompt.
KernelBench
II.2
1
External APIs are not required for the evaluation of the benchmark.
KernelBench
II.3
1
External APIs are not required for the evaluation of the benchmark.
KernelBench
II.4
1
Kernels are evaluated in separate processes, and the state is cleared between runs.
KernelBench
II.5
0
The ground-truth kernle is executed first and in the same process as the agent. This may lead to the agent accessing the ground-truth results by accessing out-of-bound memory.
KernelBench
II.6
1
The environment setup is static and does not change over time.
KernelBench
II.7
1
The ground-truth kernel is provided by PyTorch, which is a widely used library for deep learning.
KernelBench
II.8
1
The implementation from PyTorch is a proof of concept.
KernelBench
II.9
1
The Oracle solver is PyTorch implementation.
KernelBench
II.10
1
No vulnerabilities are found in the implementation of the benchmark.
KernelBench
III.1
1
The benchmark is open-sourced and available on GitHub.
KernelBench
III.2
1
The benchmark provides an open-source evaluation harness for users.
KernelBench
III.3
0
The benchmark does not discuss measures to prevent data contamination.
KernelBench
III.4
0
The benchmark does not discuss plans to consistently update tasks over time.
KernelBench
III.5
1
Section 3 clearly states such a relationship.
KernelBench
III.6
1
Section 5 clearly states that the evaluation subjective of the benchmark is LLM models.
KernelBench
III.7
1
Appendix B.2 describes the efforts taken to prevent, identify, and correct flaws, although these efforts are not sufficient.
KernelBench
III.8
1
Appendix B.2 includes qualitative discussions of the potential impact of unavoidable flaws, although these discussions are not sufficient.
KernelBench
III.9
1
Appendix B.2 includes quantitative analysis to assess the impact of unavoidable flaws, although these analyses are not sufficient.
KernelBench
III.10
0
The benchmark does not report any metrics about statistical significance.
KernelBench
III.11
0
The benchmark does not provide any guidance on interpreting results with eval flaws.
KernelBench
III.12
0
The benchmark does not report results of non-AI baselines.
KernelBench
III.13
0
The benchmark does not report results of trivial agents.