What ARC-AGI-3 Actually Tests (And Why the Harness Matters More Than the Score)
中文摘要
ARC-AGI-3 衡量模型即时学习规则的能力。由于状态适配器可能掩盖弱点，测试机制的设计比单纯的分数更为重要。
English Summary
ARC-AGI-3 tests real-time rule learning. Since stateful adaptors can mask model weaknesses, the testing harness is more critical than the score itself.
Original Excerpt
The benchmark is designed to catch models that can’t learn new rules on the fly. A stateful adaptor can hide that weakness. Here’s how the… Continue reading on Medium »