In general: there is a strong correlation between adding test-time search (or test-time training) to your model and ARC-AGI performance. There is zero correlation between adding vision as a modality and better performance. It's all about better reasoning, not at all about vision.

→ View original post on X — @fchollet