![Overview figure for Beyond [CLS]](/assets/img/pub/beyond_cls.webp)
Beyond [CLS]: Exploring the True Potential of Masked Image Modeling Representations
Why do masked image models look weak out of the box? Their attention is spread almost uniformly over patches, so the [CLS] token aggregates them poorly. Selective Aggregation makes better use of the semantic information already present in the patch tokens.
Also presented at the UniReps Workshop at NeurIPS 2025.

