Researchers have introduced ActionSplice, an inference framework for editing actions mid-generation in chunk-autoregressive video world models. The work, posted as a paper on Hugging Face, addresses a structural limitation of these models: each generated chunk is conditioned on a single action, so an action received while sampling is underway must either wait for the next chunk, condition future solver evaluations on a state produced under the previous action, or trigger a rollback that repeats already-completed evaluations.
ActionSplice reframes the problem as Counterfactual State Transport (CST). At the interruption step, a lightweight corrector transports the interrupted backbone-native representation toward the matched state that the revised action would have induced at the same solver step. Crucially, the world model and sampler remain frozen, and sampling resumes without replaying completed evaluations.
The paper describes two variants. CST-R retargets the entire active chunk, while CST-T preserves a temporal prefix and updates only the suffix. Both are evaluated across the minWM-Wan Action2V and HY-WM1.5 settings.
According to the reported results, CST-R reduces rollback-relative LPIPS by 61.5% and 75.9% relative to direct condition swapping. CST-T reduces suffix LPIPS by 56.1% and 77.5%, respectively, while delivering pixel-ready speedups of 2.73× and 1.69× over waiting. Under the HY-WorldPlay protocol, CST-R reaches a PSNR of 25.66 dB, an SSIM of 0.6902, and an LPIPS of 0.1337 against the original rollout.
The authors position ActionSplice as enabling in-flight action editing for interactive world models, a capability relevant to real-time or user-driven generation where actions may change after sampling has begun.