Just Action Models showed that future supervision improves robot control with no visual generation at inference, beating FastWAM on LIBERO-Plus at 64.5% versus 55.3% while cutting estimated online training FLOPs by 73%
AspireAI community · arXiv (Waseda / Preferred Networks)
Vision-Language-Action and world-action model communities disagree about whether robot control policies should inherit knowledge from vision-language models or from visual generators, an argument the authors say bundles three separable choices: backbone initialization, the objective applied to future observations, and whether the machinery for that future loss is kept at inference. Just Action Model, an arXiv paper submitted October 27, 2026, decouples them into a pretrained visual backbone with an action expert trained with both an action loss and a foresight loss, where the foresight loss trains current-observation features to predict frozen encodings of the recorded future through a prediction head discarded after training, so no visual generation happens in training or inference and the backbone is free. JAM policies outperform their matched FastWAM and ImageWAM counterparts on LIBERO and LIBERO-Plus and match or slightly beat them on RoboTwin; against FastWAM the success rates move from 97.0% / 55.3% / 91.2% to 98.2% / 64.5% / 91.4%. With cached foresight targets, JAM reduces estimated online training FLOPs per sample per update by 73% relative to FastWAM. The limitation is that the result is confined to simulation benchmarks and the authors' own comparisons, and the framework still requires foresight targets, which cost compute when they cannot be cached.
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.