Spent some time testing what vision language models actually understand versus what they can describe. A chessboard turned out to be a great probe because there is one correct answer for the layout (the FEN string). The models usually recognize the pie…