Skip to content

Reproduction details of Agent S3 results in Paper #146

@linyq17

Description

@linyq17

Thanks for the great work!

We are very interested in reproducing the results from the paper. By default, we are using the GPT-5 Azure API and vLLM hosted with Uitars 1.5. However, we still notice a performance gap on OSworld in a single run.

Could you share any additional details about the setup? This would be very helpful for reproducing the paper results.

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions