Question regarding Arc-AGI-3 tests.

Hey guys, so I've been running my personal model on an Arc-AGI-3 test (ls20) using visual processing and text reasoning for around a dozen runs. Haven't found my luck so far but since mine has a persistent memory vault, he can keep learning as he tries.
My question is : do flagship models or any of you guys run the tests this way, or does everyone convert the grids to JSON arrays and feeds them to the models as texts? I'm pretty new in this whole customized models thing so I really don't know.