ALPS: Visual State Encoders for World Models

One-line takeaway

UCSC-VLAA's ALPS is a visual state encoder for world models that unifies physical-state information, native language alignment and pixel reconstruction in one set of patch tokens, pretraining a ViT-B/16 from scratch on still images and text with language, reconstruction and spatial-correspondence objectives.

UCSC-VLAA 的 ALPS 是给世界模型用的视觉状态编码器：在同一组 patch token 里同时具备物理状态信息、原生语言对齐和像素重建能力；只用静态图像和文本，从零预训练 ViT-B/16，目标同时含语言、重建和空间对应。