Encoder achieves lowest state-probe error among ViT-B models in evaluation
The encoder was pretrained using still images and text with language, reconstruction, and spatial correspondence objectives, then refined through cross-view self-distillation.
In evaluation across fifteen public encoders, the encoder achieved the lowest aggregate state-probe error among ViT-B models while maintaining strong language alignment and accurate pixel reconstruction.
Sources behind this report
- Hugging Face
Research · abstract · CC0 1.0
Source published
Read by the newsroom
Updated Oct 21, 2026, 9:41 PM Central
The encoder was pretrained using still images and text with language, reconstruction, and spatial correspondence objectives, then refined through cross-view self-distillation.
In evaluation across fifteen public encoders, the encoder achieved the lowest aggregate state-probe error among ViT-B models while maintaining strong language alignment and accurate pixel reconstruction.
Sources behind this report
- Hugging Face
Research · abstract · CC0 1.0
Source published
Read by the newsroom
Updated Oct 21, 2026, 10:40 PM Central
The encoder was pretrained using still images and text with language, reconstruction, and spatial correspondence objectives, then refined through cross-view self-distillation.
In evaluation across fifteen public encoders, the encoder achieved the lowest aggregate state-probe error among ViT-B models while maintaining strong language alignment and accurate pixel reconstruction.
Sources behind this report
- Hugging Face
Research · abstract · CC0 1.0
Source published
Read by the newsroom
Report history
Source-record dates and replacement dates describe the reporting, not the order of events.
2 retained revision dates
- Revision 2Source record as of Replaced
- Revision 1Source record as of Replaced
This brief was made by AI and checked against its cited public sources. How the newsroom works