ARC-AGI-3: Learning and Planning in Unknown Environments

I am working on ARC-AGI-3, a benchmark of interactive games where an agent must infer mechanics and goals purely through interaction, under offline evaluation constraints with no external supervisor. My main focus is on developing agentic systems that combine reinforcement learning, search, and case-based memory to enable efficient decision-making and generalization, targeting state-of-the-art performance. I also work on using large language models to synthesize executable world models from interaction traces to support downstream planning.