What Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling
World action models (WAMs) predict the future alongside actions during training. Due to the heavy computation cost of video denoising, whether the future must still be generated during inference is disputed: Explicit WAMs denoise it into clean frames along with every action chunk, whereas Latent WAMs discard it entirely for acceleration. We find that latent WAMs, despite matchi
Reading in Plain English mode — toggle below to switch
World action models (WAMs) predict the future alongside actions during training. Due to the heavy computation cost of video denoising, whether the future must still be generated during inference is disputed: Explicit WAMs denoise it into clean frames along with every action chunk, whereas Latent WAMs discard it entirely for acceleration. We find that latent WAMs, despite matchi
This research highlights a critical trade-off between inference speed and generalization capabilities in World Action Models (WAMs). Understanding these nuances is crucial for ML engineers designing agents that need to perform robustly in varied environments, ensuring that optimization efforts don't inadvertently compromise core model strengths.