Rollback let robot policies rewind world models to past decisions, lifting real-world success by 10.67%, 11.11% and 11.85% on three tasks and beating average human performance in 8 hours
AspireAI community · arXiv (MIT / Shanghai AI Lab)
Rollback, an arXiv paper submitted October 26, 2026, argues that action-conditioned world models are used only as forward-only surrogates of the physical environment, leaving unused a capability no physical environment affords: rewinding to a past decision point to explore counterfactual futures. Rollback rolls back to critical decision points, branches into alternative actions and contrasts their outcomes to attribute success to individual decisions, distilling counterfactual outcomes into action-level supervision without a separately learned value model. Its training system unifies distributed world-model search with asynchronous real-robot execution in a self-reinforcing cycle, where real experience sharpens the world model and a sharper world model drives further policy improvement - a capability the authors call test-time physics grounding, achieved through a lightweight LoRA pipeline that corrects the learned dynamics against real transitions. On three real-world tasks the authors report absolute success-rate gains over the strongest baseline of 10.67%, 11.11% and 11.85% respectively, with more stable improvement throughout training, and the policy surpassing the average success rate of human players on both ball tossing and tabletop curling within 8 hours of training despite starting from teleoperated demonstrations far below human level. The honest limitation is that this is a self-reinforcing loop that depends on a world model accurate enough for its counterfactuals to be worth branching on, and the evaluation covers three tabletop tasks rather than broader deployment.
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.