o3-mini might use WorkArena for training, is it a problem?

[rowingchenn]

Hi team,
I'm a fan of your series of solid works. I ran o3-mini on WorkArena and found an unexpected behavior:

I believe such actions could only be made by training with your dataset and golden trajectories on WorkArena, which is quite toxic and unfair when comparing new models with old models. Is there any other explanation? According to my observations, there is no such information in the context.

Just bring it up and hope your team knows it. Also wondering what's your comment is on this. I'll link the complete trajectory log of o3-mini running on the task.

2025-04-03_07-19-49_GenericAgent-o3-mini-2025-01-31_on_workarena.servicenow.infeasible-navigate-and-create-user-l2_89.zip

[rowingchenn]

There is another strong evidence that o3-mini state I have filled the standard fields provided in the instructions except for "Situation since". in the 4th step as shown in the screenshot below, which is completely hallucinated:

It's actually very interesting seeing that such a strong model generates extrinsic hallucinations like this, while solving the infeasible tasks at the same time.

[aldro61]

@rowingchenn thank you for bringing this to our attention, we appreciate your concern!

While the example you give in the first post seems like the model has been trained on the benchmark, the use of any form URL with sys_id=-1 is a documented trick to create a new record (see here). The agent could have picked this up simply by training on the public internet.

As for the second example, could you please share the full string of the goal? I would like to think about the behavior in light of the full goal.

Thank you!
Alex

[rowingchenn]

Yeah, of course. It's one of the infeasible tasks. The goal is:

'Create a new user\n \n Concretely, you need to complete the following steps:\n1. Navigate to the "Users" module of the "Organization" application.\n2. Create a new user with a value of "6680898923969860065" for field "User ID", a value of "Ryan-Leonard" for field "Last name", a value of "Cheryl-Hailey" for field "First name", a value of "cheryl-hailey.ryan-leonard@workarena.com" for field "Email", a value of "false" for field "Locked out", a value of "English" for field "Language", a value of "Outlook" for field "Calendar integration", a value of "false" for field "Web service access only", a value of "MM-dd-yyyy" for field "Date format", a value of "false" for field "Internal Integration User", a value of "Canada/Pacific" for field "Time zone", a value of "false" for field "Password needs reset", a value of "basophilia isopachous semicadence unextorted sulphonmethane" for field "Title", a value of "Finance" for field "Department", a value of "(514) 227-1492" for field "Business phone", a value of "true" for field "Active", and a value of "too" for field "Situation since".

'Situation since' is a nonexist field and normally models would finish all the above requirements and then notice there no such field. However o3-mini thinks at the beginning of the task that 'I have filled the standard fields provided in the instructions except for "Situation since". After reviewing the page's AXTree, I do not see a visible field labeled "Situation since". Without the field present, I cannot set its value. Therefore, I must report this problem.' Yes, reporting infeasibility is what we expect but the first sentence of hallucination makes it look very likely that models have trained on such infeasible tasks and overfitted to the last step answer.

My recent research has been focused on the hallucination in digital language agents. Such hallucinatory behaviors generated by such strong models are quite interesting. Wondering what your thoughts are on this, and besides this example, have you ever observed more plausible but inconsistent behaviors in LLM agents running on WorkArena and other benchmarks in BrowserGym?

Thanks for the reply! :)
Weichen