ActionSplice: In-Flight Action Editing for Interactive World Models
Pardis Taghavi, Tingyu Guo, Jonas Lossner, Gaurav Pandey, Reza Langari
Texas A&M University
Published September 8, 2026 · on the feed September 14, 2026 · ▲ 10 · arXiv 2609.08230
The abstract
Chunk-autoregressive video world models typically condition each generated chunk on one action. An action received during sampling must therefore wait for the next chunk, condition future solver evaluations on a state produced under the previous action, or trigger rollback that repeats completed evaluations. We introduce ActionSplice, an inference framework that formulates this problem as Counterfactual State Transport (CST). A lightweight corrector transports the interrupted backbone-native representation toward the matched state induced by the revised action at the same solver step. The world model and sampler remain frozen, and sampling resumes without replaying completed evaluations. The retargeting variant CST*{R} updates the entire active chunk, while the temporal-splicing variant CST*{T} preserves a temporal prefix and updates only the suffix. Across minWM-Wan Action2V and HY-WM1.5, 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 providing 2.73times and 1.69times pixel-ready speedups over waiting. Under the HY-WorldPlay protocol, CST_{R} obtains a PSNR of 25.66 dB, an SSIM of 0.6902, and an LPIPS of 0.1337 against the original rollout.
The Brief
2:12A short podcast summary of the paper. Written and voiced by AI.
Transcript
ActionSplice. In-Flight Action Editing for Interactive World Models. By Pardis Taghavi and four others, from Texas A and M University. Published on the eighth of September. Its code repository is worked out: about forty source files, with a readme and a licence.
Video world models that generate chunk by chunk usually condition each chunk on a single action. That becomes awkward when a new action arrives in the middle of sampling. There are three options, each with a cost. Wait for the next chunk. Keep conditioning the remaining solver evaluations on a state produced under the previous action. Or roll back, repeating evaluations that were already complete.
ActionSplice is an inference framework that treats this as a transport problem, called Counterfactual State Transport, or CST. A lightweight corrector takes the interrupted representation, native to the backbone, and transports it toward the matched state that the revised action would have induced at the same solver step. The world model and the sampler stay frozen, and sampling resumes from there without replaying any completed evaluations.
It comes in two variants. The retargeting variant updates the entire active chunk. The temporal-splicing variant preserves a temporal prefix and updates only the suffix, so what came before the new action is left as it was.
Tests run on minWM-Wan Action2V and on HY-WM1.5. Against swapping the condition directly, retargeting reduces LPIPS measured relative to rollback by sixty-one point five percent on the first model, and seventy-five point nine percent on the second. Temporal splicing reduces suffix LPIPS by fifty-six point one and seventy-seven point five percent respectively.
Temporal splicing also delivers pixel-ready speedups over waiting: two point seven three times on one model, one point six nine times on the other. Under the HY-WorldPlay protocol, retargeting reaches a PSNR of twenty-five point six six decibels, an SSIM of zero point six nine zero two, and an LPIPS of zero point one three three seven, measured against the original rollout.
The move is to treat a late action as a correction to a state, not a reason to start over. Everything expensive stays frozen, and a small corrector carries the half-finished sample to where the new action would have taken it.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Fast video is not the same as responsive video. Here is the gap.
You are playing inside a generated video world. The model makes the video in chunks of frames, and each chunk is produced under one action, say walking forward.
The road ahead
- 1 · The problem ~7 min When you change your action halfway through a chunk, the video model has already spent work shaping that chunk for your old action.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Jump the interrupted draft straight to where a redo with the new action would be at this same step, using a small learned corrector.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Receive the new action, hand the interrupted state to a small corrector, move it to the matched state, then finish the remaining steps.
- 6 · The experiment ~6 min You’ll guess: On HY-WM1.5, by how much did the retargeting variant cut rollback-relative LPIPS compared with direct condition swapping?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.