Do you mean that sitting with the problem for days is a weakness that you should fix since you're wasting time making no progress? Or that it is a necessary practice in order to understand your weaknesses?
ARC-AGI-3 is the latest benchmark from Arc Prize. It has an interesting departure in that each puzzle is now interactive, like a mini RL task. Players are not told the goal, requiring the agent to explore and model the puzzle to solve it.
Is anyone giving this an attempt? May or may not be LLM-powered. Would love to hear your approaches.