Have you tried a grid representation without image_url? I wonder how important was the image in solving the puzzles. Ryan complained about GPT4o's vision on grids. LLMs would be better at ARC if they had better vision. Any thought?
Have you tried a grid representation without image_url? I wonder how important was the image in solving the puzzles.
Ryan complained about GPT4o's vision on grids. LLMs would be better at ARC if they had better vision. Any thought?