Paper Link: (https://arxiv.org/pdf/2511.14761)

This paper had me hooked throughout this weekend. One of the authors is the OG Kaiming He, so I had to dig in and implement (plus the official repo is not out yet and I could not wait).

Computer Vision is so back! The authors took a contrarian appraoch to solving the ARC challenge. Although, calling it contrarian feels a bit weird because ARC is fundamentally a Vision problem, yet LLM based solutions dominate the leaderboard.

If you don’t know ARC yet, it’s one of the most interesting benchmarks in AI. Each task is a tiny colored grid puzzle. You’re given a few (input → output) examples and asked to infer the transformation rule and apply it to a held-out input. The rules are deliberately “human-ish”: mirror this pattern, move the shape up, fill the enclosed region, copy the blue object everywhere there’s green, etc.

Most approaches treat ARC as:

a language / program synthesis problem : search over programs, or let an LLM emit a symbolic description of the rule.

It’s really refreshing to see ARC being modelled as an image2image translation problem & getting an amazing 60.4% on ARC test set, with just an 18M ViT that was trained on 400 tasks from ARC train set.

##