webarena
webarena copied to clipboard
Hard to read the trace and to accomplish the task regarding map domain
Hi, there. I am working on some map tasks and have found it hard to accomplish them even by myself. The usage of the OpenStreetMap is quite different from the usage of Google Maps.
Specifically, for task 57: Tell me the closest restaurant(s) to university center at Carnegie Mellon University. I tried the keywords: "restaurants near university center at Carnegie Mellon University", "restaurants near Carnegie Mellon University", "restaurants near CMU", none of them shows an effective answer. Only "restaurants near university center" could lead to some results. The showed-up restaurants don't seem to be ranked according to distance. How are they ranked? Moreover, the eval_type of 57 is "must_include" with 7 restaurants' names. I found it really hard to finish it as human.
Another example is for task 36: "social security administration in Pittsburgh" shows no results but "social security administration Pittsburgh" does. So I wonder if there are some rules about using the OpenStreetMap?
I looked at the trace of 58: Tell me the closest cafe(s) to CMU Hunt library. The human actions include typing the keyword "Cafe near Hunt library", and then playing around with the results (zooming in and out). How is it to be judged the second result is the final answer? Maybe according to the visual information to look at the location of CMU campus? If so, is it impossible for a text-prompt-agent to solve the problem?
I attached the config file links for your convenience. Task 36 Task 57 Task 58
Thanks!