Giving a world action model harder visual noise where the scene is moving, and easier noise where it is static, lifted robot success from 38.8 to 47.5 percent
AspireAI community · arXiv (cs.CV, cs.AI)
World Action Models borrow pretrained video generators to predict both what a robot will see next and what it should do. The standard visual objective treats every pixel as equally worth predicting, which is odd for control: a static wall teaches a robot nothing, while a moving arm and a grasped object carry all the signal. Motion-Guided Asymmetric Prediction (MoGAP) changes that allocation instead of adding a teacher or a new modality.
MoGAP corrupts dynamic regions of the video harder than static ones and asks the shared video-diffusion backbone to reconstruct them using the static regions as context. The motion cues come straight from the training videos, and the auxiliary prediction path exists only during training. On GR-1 Tabletop, the paper reports success rates rising from 38.8 to 47.5 percent for Fast-WAM (Wan2.2-5B) and from 52.3 to 62.3 percent for Cosmos3-Nano, outperforming competitive VLA and WAM baselines as well as representation-alignment methods that use external teachers. On a real-world OpenArm humanoid, the trained policy responds better to task-relevant perturbations such as a target object being relocated mid-execution, even without demonstrations of such perturbations.
The caveats are scope rather than rigour. Every number is the authors' own run on GR-1 Tabletop, RoboCasa-Kitchen and one OpenArm humanoid; the real-world result is reported qualitatively in the text with success rates and standard errors living in a figure rather than a headline figure. Deriving motion cues from the training videos assumes those videos contain the motion signal the method needs, and the asymmetric noise path is training-only, so nothing changes at inference time - which is exactly why it is cheap.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.