level: technical
recent progress on the arc-agi-1 benchmark has come from two main approaches: heavy test-time compute with frontier models or fine-tuning small models on arc data. this work explores a third path using an open-weight model, deepseek v3.2, in non-thinking mode under a strict budget and without any arc-specific fine-tuning. the focus is on what can be achieved through architecture alone by building agentic harnesses that explicitly separate pattern discovery and program synthesis.
the first component is an explorer-definer pipeline that splits the process into two stages. one stage discovers patterns in the input-output examples, and the other synthesizes executable transformations based on those patterns. this separation allows the system to handle abstract reasoning tasks more systematically. the second component is a reflective orchestrator that adds autonomous exploration of new transformations when initial hypotheses fail, improving the system's ability to generalize across different tasks.
the approach shows that competitive performance on arc-agi-1 is possible without relying on massive compute or benchmark-specific training. by structuring the reasoning process into distinct stages and adding reflective capabilities, the system can solve complex abstract reasoning problems efficiently. this opens up possibilities for cost-effective ai systems that can generalize better from limited data, making advanced reasoning more accessible.
why it matters: this method reduces the cost and complexity of building ai systems that can reason abstractly, making them more practical for real-world applications where data and compute are limited.