A controlled ablation found future-video prediction lifts robot task success by up to 72.91 points, but the gains do not reliably survive distribution shift
AspireAI community · arXiv
World Action Models get their robot policies by also training them to predict what the camera will see next, and the headline claim for that design is better task success. This paper takes the claim apart properly: it ablates the future-video prediction loss inside two model families while holding architecture, initialisation, training data and training steps fixed, so the only thing that changes is whether the model is asked to predict the future.
On clean conditions and tasks the robot has already seen, that supervision is worth 24.38 and 72.91 percentage points of success for the two models. Trajectory clustering shows the gains are not just a shift in which behaviours the robot performs more often - success rises inside trajectory clusters whose frequency barely moves. But the paper's most useful finding is the limitation: under unseen tasks and environmental randomisation, the advantage does not hold consistently, and both models score low. A lightweight Transformer predictor used only during training also lifted pi0.5 from 82.94% to 89.19%, suggesting the benefit does not require a large pretrained video diffusion backbone.
The caveat is that these are the authors' own numbers on two model families and one robot platform, and the large figures describe the easy case rather than deployment conditions - which is precisely the distinction the paper is making.
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.