Physix Frontier · News Briefing Card (QbitAI · Oct 1, 2026)

He Kaiming team solves ARC with ImageNet-pretrained vision model

KEY FACTS

He Kaiming's team proposed NAT-ARC, a pure vision approach to ARC abstract reasoning tasks that does not rely on large language models.
The method first uses an ImageNet MAE pretrained encoder, then trains on the ARC training set, and applies LoRA fine-tuning per task at test time.
NAT-ARC scores 63.4% pass@2 on ARC-1 as a single model and reaches 70.2% with an ensemble.
Compared with the specialized LLM approach The ARChitects at 71.6%, NAT-ARC uses about one quarter of the parameters.
The paper found that after pretraining the model's scaling curve turns positive, whereas previously without pretraining, larger models performed worse.

KEY DATA

63.4%Single-model ARC-1 pass@2

70.2%Ensemble ARC-1 pass@2

0.6BSingle-model parameter count

about 2BTotal ensemble parameter count

PHYSIX OBSERVATION

The pure vision route uses pretraining to patch up its scaling weakness; with far fewer parameters than LLMs it approaches specialized systems, showing that abstract reasoning does not necessarily depend on language. For the industry, this suggests visual priors can transfer to symbolic tasks, that ARC solutions may no longer be monopolized by LLMs, and that small models plus pretraining deserve reevaluation.

Source: QbitAI report