AlayaWorld v1.1 Switches to a Streaming 3D Point Cache, Raising Its Long-Horizon World-Model Consistency Score to 89.5

Alaya Lab has released the AlayaWorld v1.1 technical report. Rather than scaling up the 15B-parameter model, the update focuses on correcting the representational mismatch between conditioning signals and generated latents. The original chunk-wise autoregressive backbone and training data remain unchanged. The new version replaces its depth-warping-based spatial memory with a streaming 3D point cache. After each chunk is generated, per-pixel 3D points are registered in a persistent cache and then re-rendered from the planned camera position for the next chunk. Camera trajectories are scale-aligned using existing displacement, while the intrinsics are fitted from the point map and then fixed.

The second major change makes conditioning inputs and outputs follow the same causal VAE temporal rules. Instead of encoding a static image as a single frame, the system places it in a nine-frame window to obtain a latent with temporal context. The same method is used for transitions between chunks. Temporal memory is first aligned over 25 frames in pixel space and then encoded into exactly four latents, preventing the previous version’s off-by-one boundary from leaking into the context. Memory dropout has also changed: instead of retaining token positions with all-zero tokens, the tokens are removed entirely, ensuring that training sees the same memory-free sequence used during initial inference.

The new version also removes the separate camera AdaLN branch, expressing camera control entirely through re-rendered output from the 3D cache at the target viewpoint. This allows parallax, occlusion, and scale to enter the model through the same visual latent representation, while valid regions can directly use FlashAttention.

Across WBench’s 158 navigation cases, AlayaWorld achieved an average consistency score of 89.5, the highest among the nine systems tested. It also ranked first in background, perspective, subject, and geometric consistency. However, its navigation score of 80.0 was lower than HY-World 1.5’s 87.5, while its average physics score of 63.1 trailed the leading score of 69.1. This shows that preserving a scene is not the same as correctly simulating causality. More importantly, the GitHub repository still lists the “improved weights” and training code as incomplete, while the public quick-start guide continues to describe the older DA3 and AdaLN pipeline. Engineers should wait until the updated weights, code, and ablation results are aligned before determining whether this architectural update can be reproduced independently.