Token-World: World Modeling in Vision-Language Model Token Space for Robot Manipulation

Vision-language-action (VLA) models are a widely adopted paradigm for embodied policies. They excel at efficient closed-loop control but do not explicitly model how physical scenes evolve as a task unfolds. Recently emerging world-action models (WAMs) leverage pretrained video world models to capture spatiotemporal evolution, yet retaining future generation or a large video backbone in the control loop substantially increases inference cost. We introduce World Tokens, an embodied policy architecture built around a World Adapter that bridges visual-language understanding, world-dynamics modeling, and action generation. It uses world modeling during training to enhance the action policy while preserving efficient deployment. Specifically, the World Adapter transforms VLM features into a fixed set of world tokens, which condition a jointly fine-tuned future-video denoiser and simultaneously serve as the action expert's sole visual-language context. This shared conditioning allows gradients from future-video denoising to directly shape the representation used for action prediction, while exclusive routing prevents the policy from bypassing that representation. At deployment, the world-model branch is removed, leaving only the VLM, World Adapter, and action expert, with no online video-model inference. With a 2B backbone and no embodied action pretraining, World Tokens is highly competitive on LIBERO, attains the best reported averages on SIMPLER, substantially improves real-world R1 Pro success over a matched action-only baseline, and generates each action chunk at VLA-level latency.

Vision-language-action policies inherit both capabilities and input representations from pretrained vision-language models, showing great potential for robotic manipulation across diverse industrial settings. As these policies increasingly use interaction history, organizing the representation of historical observations determines how experiences enter temporal context and how relationships across time are modeled, which is a generally ignored challenge in previous works. In this work, we believe solving this challenge requires an architectural reconstruction and propose \textbf{WorldToken}. WorldToken encodes each timestep's observations into one world token, processes the resulting history with a causal Transformer, and generates action chunks with a diffusion action head. Unified token enables long horizon tasks while relieving the memory requirement of the temporal backbone, hence improving performance. This design also offers high interpretability and allows advances in language modeling, such as pre-training and scaling, to be transferred to robot interaction policies. In RoboCasa experiments, WorldToken successfully handles most tasks with 85M parameters and achieves 59.4% mean closed-loop success close to $π_{0.5}$ with 3.35B parameters. On the memory benchmark RMBench Blocks Ranking, WorldToken can reach the context of two minutes and achieves a success rate of 95%. In addition, holdout action RMSE is well described by power-law fits, and its closed-loop success rate improves consistently with increasing training data size in a study involving approximately 350,000 closed-loop evaluation episodes across 50 trained policies on RoboCasa, showing its scaling potential. We also conduct experiments to analyze information preservation and history use in WorldToken, providing empirical grounding for future work.

Vision-Language-Action (VLA) models map observations to actions with no objective that accounts for how the world responds, so their robustness is bounded primarily by data coverage. World models carry precisely that missing objective and are better grounded for it, yet rolling the future forward costs seconds per decision and rules them out of the control loop. We show the two can be separated. What a world model knows about physical scenes lives in its internal features; generating the future is merely the objective that produced them, so the grounding can be inherited while the generative machinery is left behind. We add one feature-alignment term to ordinary VLA training: a frozen world model is run over the training frames once and cached, and the student learns to agree with that cache. No teacher is loaded during training, the projector is discarded after it, and the deployed policy is identical to the undistilled baseline, running in 32ms and 1.86GB on a consumer RTX5090, so every gain is attributable to the representation rather than to added capacity or test-time compute. A 0.8B student reaches 97.9% on LIBERO, improves from 48.2% to 50.5% on RoboCasa-GR1 humanoid manipulation, and the same objective carries over to real hardware, on both a single-arm and a bimanual platform. The gain survives changes of student scale, backbone, alignment layer, and teacher, indicating a broad representational prior rather than a fragile alignment between two particular networks. Project page: https://thaw-vla.trung-dt.com/.