Video World Models · 2026-11-03
Old Is Not False: Valid-Time Memory for Interactive Video World Models
Ziming Wang, Chuyue Gong, Hongliang Cao, Jingmin Zhu, Chutian Wang, Yuheng He, Xiuyu Li
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors identify an undo failure mode in interactive video world models: when an agent restores a region to an earlier state, memories that rely on capture-time ordering can make generators draw outdated content. ReDate keeps all frames but assigns tokens a valid time derived from the agent’s action log, so a reverted region is re-dated as having “just reached” its state. The authors report large gains on FAR and HY-WorldPlay reverted steps, and provide theory and a calibrated state metric.
Why it matters
Interactive video world models must use an external memory of past frames. The paper argues that “old” should not be treated as “false” after undos: capture time (transaction time) differs from the valid time that determines what content should be trusted. ReDate formalizes valid-time memory, proves when latest-timestamp reading works, and improves reverted-step rendering accuracy on reported benchmarks.
Method
- Hold frames/weights/noise fixed and probe how timestamp order affects FAR and other generators, showing a soft latest-timestamp/last-writer-wins behavior after re-timestamping.
- Introduce ReDate: keep every frame, but re-date retrieved tokens using valid time derived from the agent action log; the frozen generator then interprets a restored region as just updated without training.
- Provide theory for when latest-timestamp reading is correct (valid time conditions) and propose a calibrated state metric to test memory after undos.
Limitation
The authors state that “fourfold stretching lowers [the outdated share] to 0.05 and sends 0.62 of the outputs below the quality floor, outside the span of timestamps FAR was trained on.”
Abstract (from arXiv)
Interactive video world models draw each frame from a memory bank of earlier frames stamped with their capture times. With frames, weights and noise fixed, reversing the timestamps of two views flips the state that generators of two families draw: they take the newest timestamp as the truth. That reading is right as long as states never return. After an agent undoes an action, the newest view of a region is outdated and the only view of the present is older; memories for changing worlds delete that older view, and memories that keep it hand the generator the past. Temporal databases separate when a record is stored from when its content holds, its valid time. ReDate keeps every frame and dates each retrieved token by its valid time, read from the agent's action log. The frozen generator then reads a reverted region as one that has just reached a new state, so ReDate needs no training. We prove that trusting the newest timestamp is right exactly when every current view is dated after every outdated one, that without a view taken after the return only the action log tells a reverted region from an up-to-date one, and that a generator able to draw from its own prediction gains more from re-dating than from deletion when the moved tokens keep little weight. After an undo, ReDate raises the share of FAR's steps showing the current state from 0.36 to 0.62 and HY-WorldPlay's from 0.003 to 0.19. We also propose a calibrated state metric to test any memory after undos.
Related papers
- WorldMemBench: A Comprehensive Benchmark for Memory of Video World Models
- FutureWorld: A Plug-and-Play Causal Next-State Predictor for Interactive Video World Exploration
- GROME: GROunding Factorized MEmory into Video World Models
- World-T3: Test-Time Training as Persistent Memory for Real-Time Video World Models
- StateCraft: Augmenting Video World Models with Abstract-State Reasoning
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-11-03 · Indexed by Pub-AI: 2026-11-03 UTC
Old Is Not False: Valid-Time Memory for Interactive Video World Models
First announced by arXiv: 2026-11-03 · Original record ↗
- arxiv
- 2611.00758
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-11-03 · First detected: 2026-11-03 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.00758v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
Interactive video world models draw each frame from a memory bank of earlier frames stamped with their capture times. With frames, weights and noise fixed, reversing the timestamps of two views flips the state that generators of two families draw: they take the newest timestamp as the truth. That reading is right as long as states never return. After an agent undoes an action, the newest view of a region is outdated and the only view of the present is older; memories for changing worlds delete that older view, and memories that keep it hand the generator the past. Temporal databases separate when a record is stored from when its content holds, its valid time. ReDate keeps every frame and dates each retrieved token by its valid time, read from the agent's action log. The frozen generator then reads a reverted region as one that has just reached a new state, so ReDate needs no training. We prove that trusting the newest timestamp is right exactly when every current view is dated after every outdated one, that without a view taken after the return only the action log tells a reverted region from an up-to-date one, and that a generator able to draw from its own prediction gains more from re-dating than from deletion when the moved tokens keep little weight. After an undo, ReDate raises the share of FAR's steps showing the current state from 0.36 to 0.62 and HY-WorldPlay's from 0.003 to 0.19. We also propose a calibrated state metric to test any memory after undos.
