ThinkWorld turned robot actions into Chain-of-Dynamics tokens so unlabeled human video could train a VLA policy, hitting 97.3% on LIBERO and 70.5% under robot initial-state changes
AspireAI community · arXiv (Seoul National University / NYU)
ThinkWorld, an arXiv paper submitted October 27, 2026, targets a fundamental trade-off in robot learning: action-labeled demonstrations are scarce and narrow, while human video is plentiful but lacks explicit action grounding. ThinkWorld represents each action thought as a Chain-of-Dynamics, a sequence of discrete motion tokens capturing the intermediate dynamics between the current world and a future state, used as a common action representation across embodiments. Chain-of-Dynamics is initialized from action-labeled demonstrations, then the model predicts the future each token induces and reinforces those whose predictions better match the observed outcome, which lets videos without action annotations supply direct supervision from observed state transitions. The reported results are 97.3% average success on LIBERO and 55.6% on LIBERO-PRO, with 31.9% and 60.9% under task and environment shifts, beating VLA-JEPA by 11.1 and 19.9 percentage points; zero-shot LIBERO-Plus reaches 71.7%, 1.8 points above DreamVLA, and 70.5% on robot initial-state changes against 40.8% for the next-best method. On real robots it succeeded in all 60 in-distribution trials and 52 of 60 out-of-distribution ones, with macro and micro success rates of 94.4% and 93.3% against 62.2% and 60.0% for the strongest baseline MolmoAct2, and a controlled pretraining study showed success rising monotonically from 45.4% for a robot-only baseline to 47.8% at a 50% human-video fraction. The limitation is that the human-video gain in that controlled study is modest, about 2.4 points, and the headline numbers are largely on the LIBERO family of simulation benchmarks.
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.