signal量子位2026-10-02
He Kaiming Team's New Work: Learning ARC Challenge by Watching Cat Pictures
NAT-ARC is a pure vision solution for ARC that uses ImageNet MAE pretraining without LLM. It achieves 63.4% pass@2 with the best single model and 70.2% with ensemble, approaching specialized LLM systems. The pretraining unlocks scaling for visual ARC methods, as larger models improve performance instead of overfitting.
- for who
- AI researchers working on abstract reasoning and visual pretraining
- why now
- He Kaiming's team just released NAT-ARC, proving ImageNet pretraining transforms ARC reasoning.
- what changes
- Pure vision ARC models can now leverage pretraining to avoid overfitting and scale effectively with model size.
- to do
- Adopt the NAT-ARC pipeline: initialize the encoder with a public MAE checkpoint, then fine-tune on ARC with LoRA at test time.
key points
- NAT-ARC uses ImageNet MAE pretraining to solve ARC without language models
- Best single model scores 63.4% pass@2, ensemble reaches 70.2% with 2B params
- Pretraining enables scaling: larger models improve, no more overfitting
#nat-arc#arc reasoning#visual pretraining#he kaiming team
score
score 8 out of 10. 0-10: how dense the facts are, multiplied by how much you can do with them after reading. 8+ means the topic's evidence bar is met: benchmarks and availability for a new model, amount and investors for a funding round, revenue figures for a solo-money story. Below 5 an item does not enter the digest. A press release scores 3 or less, a reprint loses 2, anything older than 14 days loses 1, a headline that misleads loses 3.
read the source