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.
Is anyone giving this an attempt? May or may not be LLM-powered. Would love to hear your approaches.