Kaiming He's team tackles ARC with an ImageNet-trained encoder

On October 1, 2026, the Chinese tech outlet QbitAI reported that a team led by He Kaiming has released a new work that tackles the ARC challenge with an encoder trained on ImageNet. The outlet's headline sums up the idea as learning the ARC challenge by watching cat clips, and the summary and quoted excerpt it provides say the approach relies on an encoder trained on ImageNet, with training data drawn from ImageNet.

The report gives no benchmark scores, no model architecture, no parameter counts and no training details beyond that. Nothing further about the method is described. As of this report, the public record stops at the release itself and at the fact that the method depends on an ImageNet-trained encoder for ARC: there is no follow-up, no comparison against other approaches, and no third-party replication or assessment. Because only one item of coverage exists so far, there are no conflicting claims between sources to reconcile.