🚀 Your VAR Model is Secretly an Efficient and Explainable Generative Classifier — [ICLR 2026]

[Yi-Chung-Chen]

Thanks to the great work on VAR! 🙌
Building upon it, we show that VAR models can be repurposed as generative classifiers — without any additional training.

We introduce A-VARC and its enhanced variant A-VARC+, which leverage VAR's tractable likelihoods for image classification. This unlocks two compelling capabilities out of the box:

• 🔍 Explainability — token-wise mutual information provides fine-grained visual explanations of classification decisions
• 📈 Class-incremental learning — new classes can be added without any replay data

Compared to diffusion-based generative classifiers, VAR-based classifiers are significantly faster at inference time thanks to the tractable likelihood structure.

This work has been accepted to ICLR 2026.

📄 Paper: https://arxiv.org/abs/2510.12060
🔗 GitHub: https://github.com/Yi-Chung-Chen/A-VARC