Efficient Inference for World Action Models

World action models predict actions together with how the world may change, and efficient inference studies how to run these models within real-time and memory limits. Approaches include using compact latent, motion, or object representations instead of dense video, distilling models to generate actions in few steps, and reusing cached computations across control cycles. Systems also use pruning, quantization, asynchronous streaming, and adaptive compute, with evaluation on robot manipulation benchmarks such as LIBERO and RoboTwin as well as real robots and navigation tasks. The central tension is that future prediction can aid robustness and long-horizon planning, but generating or reusing it adds cost and can introduce stale or inaccurate information.