# Alaya-EVOKE：从线性扩展监督到无尽世界 - 来源：HuggingFace Daily Papers（社区热门论文） - 发布时间：2026-08-13 08:00 - AIHOT 分数：66 - AIHOT 链接：https://aihot.virxact.com/items/cmssd3ass03ttrod0dbxttc96 - 原文链接：https://arxiv.org/abs/2608.13546 ## AI 摘要 Alaya Lab 等机构提出交互式世界模型 Evoke，通过外部相机索引世界状态库保持有界上下文，并重新设计教师模型以支持长时程监督。其稀疏注意力机制实现激活内存和计算量线性增长，30 秒长时程分布匹配目标使三步学生模型无需 CFG 即可抵抗内容漂移。 ## 正文 1]MoE Key Lab of BIPC, USTC 2]Shanghai Innovation Institute 3]Alaya Lab \contribution[*]Equal contribution, work done during internship at Alaya Lab \contribution[†]Project lead \contribution[‡]Corresponding author \correspondence\projecthttps://evoke-world.github.io/Evoke/ \codehttps://github.com/SII-YuanyangYin/Evoke Alaya-EVOKE: From Linear-Scaling Supervision to Endless World Yuanyang Yin Gongxuan Wang Yifan Zhan Chuanhao Li Feng Zhao Affiliation: Affiliation: Affiliation: yyyin@mail.ustc.edu.cn August 13, 2026 Abstract Interactive world models must simultaneously support persistent memory, responsive user interaction, and long-horizon generation, yet these requirements place conflicting demands on the underlying model. Maintaining history in the denoiser context or key-value cache incurs growing cost over time, forcing a trade-off between session length and retained memory, while low-latency interaction typically relies on few-step generation whose capabilities are ultimately bounded by its teacher. Evoke addresses both limitations by externalizing persistent world state and redesigning the teacher for long-horizon interactive generation. Scene geometry is maintained in an external, camera-indexed world state bank, from which only information relevant to the current view is retrieved, keeping the denoiser context bounded as the session grows. Rather than treating the teacher as a fixed high-quality generator, the teacher is explicitly designed for long-horizon supervision. Its sparse attention scheme combines chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state, yielding linear growth in activation memory and computation while enabling supervision over long temporal horizons. Such supervision exposes content drift that remains locally plausible within short windows, while per-chunk conditioning enables prompt changes and event control throughout the sequence. A 30-second long-horizon distribution-matching objective, applied under self-forced rollouts, transfers both capabilities to a three-step student that uses no classifier-free guidance (CFG), improving resistance to long-term content drift while preserving responsive conditioning. With bounded context and recurrent external memory, Evoke supports open-ended, continuously evolving generation; on a single H200 at 384×640, each 1.5 s chunk is generated in 2.11 s. As a three-step world model, Evoke achieves state-of-the-art performance on WBench while remaining competitive in visual quality on VBench-Long and VBench-2.0. 1 Introduction Figure 1: Two hours of uninterrupted generation. Representative two-hour rollouts under continuous camera control, generated in three steps per chunk without classifier-free guidance. Each row shows six frames sampled uniformly over the full session, illustrating sustained coherence over hour-scale generation. Interactive world models must jointly support persistent memory, responsive interaction, and long-horizon generation (Bruce et al. 2024; Ball et al. 2025; He et al. 2025b; Sun et al. 2025; Mao et al. 2025). A revisited scene should remain consistent with its earlier appearance; camera motions and text instructions should take effect promptly; and a session should extend from seconds to minutes or even hours. While substantial progress has been made on each capability in isolation, combining all three remains challenging. Their conjunction exposes a systems-level question that model scaling alone does not resolve (Yin et al. 2025b): how to maintain persistent world state without an ever-growing context, and how to equip a low-latency generator with the capabilities required for long-horizon interaction. Current interactive world models typically place two demanding requirements on the denoiser: preserving persistent world state and generating responses at interactive latency. When past observations are retained as additional context frames or an accumulated key-value cache (Chen et al. 2024; Zhang and Agrawala 2025; Li et al. 2025a; Yang et al. 2025), the cost of each denoising step grows with the session history. Windowing and cache eviction can bound this cost only by discarding information (Cui et al. 2025), while retrieval and streaming conditioning still operate under a finite denoiser-side context budget (Hong et al. 2025; Wang et al. 2026; Feng et al. 2025). Interactive generation introduces a different constraint: low latency requires inference in only a few denoising steps, which is typically achieved by distillation from a slower teacher (Yin et al. 2024b; Yin et al. 2024a; Yin et al. 2025a; Huang et al. 2025a). The resulting student is therefore constrained by the temporal horizon and conditioning capabilities represented in its teacher supervision, limiting both long-term consistency and responsiveness to changes introduced during a session. These limitations suggest that persistent state and interactive capabilities need not both be carried by the deployed denoiser. Evoke decouples them: persistent spatial state is maintained in an external world state bank, while the teacher is explicitly designed to provide the long-horizon and dynamically conditioned supervision required by the few-step student. This separation forms the central design principle of Evoke and enables bounded-cost, long-running interaction. For persistent world state, Evoke uses an explicit, bounded world state bank indexed by camera pose. Rather than retaining the observed history within the denoiser through context frames, key-value caches, or learned retrieval, the store records scene geometry externally and queries it directly from the current viewpoint. Previously observed surfaces that re-enter the camera frustum are rendered into the current view and provided to the generator as pixel-space conditioning, extending warp-based history conditioning (Wang and He 2026) from short-range frame correspondences to persistent geometric memory. As a result, the denoiser operates with a fixed number of history and rendered memory frames, keeping its token budget independent of session duration. The world state bank is itself bounded and supports explicit read, write, and eviction; the current implementation retains a finite temporal window of geometry, providing constant-cost recall within the retained coverage rather than indefinite memory of every previously visited location. The world state bank follows a geometry-based construction similar to prior work (Wang and He 2026; Ren et al. 2025; Shen et al. 2026; Wu et al. 2026): generated observations are lifted into scene geometry and later rendered from the current camera pose to provide view-aligned conditioning. This explicit geometric representation keeps persistent state outside the denoiser while preserving direct access to previously observed regions. A distilled student can acquire only capabilities expressed by its teacher supervision. Yet distillation is commonly treated primarily as an acceleration mechanism (Yin et al. 2024b; Yin et al. 2024a; Lin et al. 2025b; Yin et al. 2025a), leaving key teacher properties largely fixed: attention may remain bidirectional and quadratic, training clips short, and text conditioning constant throughout a sequence. Recent work has examined the mismatch between bidirectional teachers and causal students (Zhu et al. 2026). Evoke focuses on two complementary teacher-side design variables: supervision horizon and conditioning schedule. Short-horizon supervision provides no joint constraint over distant moments, while a single global prompt does not expose how newly introduced instructions should affect an ongoing rollout. The supervision horizon is particularly important because different failure modes emerge at different temporal scales. Degradation drift, such as exposure growth, saturation shifts, and texture degradation, alters local statistics and can therefore be detected within relatively short windows. Existing long-video objectives can substantially improve such local stability (Cui et al. 2025; Liu et al. 2025), but this improvement may coincide with reduced temporal dynamics or diversity, where suppressing variation also makes short-window statistics easier to preserve. Content drift presents a different challenge: scene identity, object appearance, or spatial layout may evolve gradually while every local window remains individually plausible. Such inconsistencies become apparent only when sufficiently distant moments are compared and therefore cannot be resolved simply by enforcing stronger local stability. This distinction makes supervision horizon an explicit teacher-side design variable: extending the teacher horizon exposes long-range inconsistencies while retaining the dynamics required for an evolving world. Per-chunk conditioning complements long-horizon supervision by exposing the teacher to prompt changes and newly introduced events within the same rollout. Evoke therefore designs the teacher jointly for long-range consistency and dynamic conditioning, providing both capabilities for subsequent transfer to the few-step student. Whether the benefit of increasing supervision horizon eventually saturates is evaluated empirically in Section 4.2. These requirements lead to a redesigned Evoke Teacher for long-horizon supervision. The teacher adopts chunk-wise sparse attention (Jia et al. 2025; Cai et al. 2025), where each chunk accesses bounded local context, a small set of distant frames, and a linear-attention global state (Yang et al. 2023; Team et al. 2025; Chen et al. 2025b), reducing attention growth from quadratic to linear with sequence length. The same chunk structure assigns an independent text condition to each chunk, allowing instruction changes to be represented within a single rollout. A 30-second full-window distribution-matching objective (Yin et al. 2024b; Yin et al. 2024a), applied to self-forced rollouts (Huang et al. 2025a), jointly transfers long-range supervision and dynamic conditioning to a three-step, CFG-free (Ho and Salimans 2022) student. At inference time, the resulting student supports scheduled prompt changes during an ongoing session, enabling elements to be introduced or withdrawn on demand, a capability we refer to as evocation (Fig. 2). Figure 2: Timed prompt switching modifies the sky while preserving anchored scene structure. Using the same input image, camera trajectory, and three-step CFG-free student, rows differ only in text. Rows 1 to 3 introduce a fireball, hot-air balloon, and fireworks, respectively, at chunk 2 and remove them at chunk 4 (amber borders denote active chunks); the objects appear and disappear while the castle, jungle, and dragon remain consistent. Row 4 uses a static castle prompt with no evocation. Open-ended generation additionally requires the recurrent process itself to remain independent of elapsed time. A rollout represented on a continuously growing positional axis (Su et al. 2024) may eventually exceed the positional range encountered during training (Cui et al. 2025). Evoke instead reuses the same local positional layout for every generated chunk, while persistent world state is accessed through camera pose rather than temporal position. Session duration therefore increases only the number of recurrent generation steps, without expanding the positional range or denoiser context of any individual call. The three-step student generates a 1.5 s chunk in 2.11 s on a single H200 at 384×640, with the world state bank enabled. Figure 1 shows seven uninterrupted two-hour rollouts under continuous camera control, demonstrating coherent hour-scale generation. Evoke further achieves state-of-the-art performance on WBench (Ying et al. 2026) while remaining competitive with many-step systems on VBench-2.0 (Zheng et al. 2025) and VBench-Long (Huang et al. 2024; Huang et al. 2025b), despite using only three sampling steps and no classifier-free guidance. Quantitative long-session evaluation is reported separately in Sec. 4.2, together with ablations on supervision horizon and conditioning granularity. These experiments isolate the two teacher-side design variables, measuring how long-horizon supervision affects content drift and how per-chunk conditioning governs mid-session responsiveness. Our contributions can be summarized as follows. (i) We formulate interactive world generation as a bounded recurrent process that decouples persistent world state from the denoiser: scene geometry is maintained in an external, camera-indexed world state bank, keeping the denoiser context and positional range independent of session duration. (ii) We redesign the teacher for long-horizon interactive supervision through chunk-wise sparse attention and per-chunk conditioning, enabling efficient supervision over long temporal horizons while exposing both long-range content drift and mid-sequence instruction changes. (iii) We transfer these capabilities to a three-step, CFG-free student using a 30-second long-horizon distribution-matching objective under self-forced rollouts, yielding bounded-cost, open-ended generation with strong long-horizon consistency and responsive mid-session control. 2 Related work Interactive video world models (Bruce et al. 2024; Ball et al. 2025; Valevski et al. 2024; Zhang et al. 2025; NVIDIA 2025; Bar et al. 2025; Decart 2026) differ fundamentally in how they represent and retain information across a running session. One line of work maintains history within the denoiser, using additional context frames or an expanding key-value cache (Chen et al. 2024; Zhang and Agrawala 2025; Li et al. 2025a), and controls the resulting growth in per-step cost through windowing, cache eviction, retrieval, or streaming video-to-video conditioning (Cui et al. 2025; Liu et al. 2025; Hong et al. 2025; Wang et al. 2026; Feng et al. 2025); several systems have demonstrated substantially extended generation horizons (Ball et al. 2025; He et al. 2025b; Sun et al. 2025; Yang et al. 2025; Chen et al. 2025a; Dalal et al. 2025; Li et al. 2025b). A second line externalizes part of this history through geometry (Wang and He 2026; Ren et al. 2025; Shen et al. 2026; Wu et al. 2026): previously generated observations are lifted into a scene representation (Mildenhall et al. 2021; Kerbl et al. 2023), using monocular depth or feed-forward geometry estimation (Lin et al. 2025a; Wang et al. 2025), and rendered into the target view, allowing revisited content to re-enter the generator as view-aligned pixel conditioning rather than as an ever-growing token history. A third line targets interactive inference through few-step distillation, building on few-step samplers developed for image diffusion (Salimans and Ho 2022; Luo et al. 2023; Song et al. 2023). Distribution-matching objectives (Yin et al. 2024b; Yin et al. 2024a; Yin et al. 2025a) and self-forcing (Huang et al. 2025a) compress slower teachers into low-step students, alongside adversarial post-training (Lin et al. 2025b), while recent work further examines the mismatch between bidirectional teacher supervision and causal student inference (Zhu et al. 2026). Evoke builds on the latter two directions while shifting the design focus from individual mechanisms to where persistent state and transferable capabilities reside. Following geometry-based conditioning (Wang and He 2026; Ren et al. 2025), Evoke renders previously observed scene content into the current view; however, the underlying world state is maintained in an explicit external world state bank with bounded read, write, and eviction, keeping the denoiser context independent of session duration. On the distillation side, Evoke adopts distribution matching (Yin et al. 2024b; Huang et al. 2025a) but treats the teacher itself as a design variable rather than merely a source of high-quality targets. Beyond attention causality (Zhu et al. 2026), the teacher is explicitly structured around two properties required for long-horizon interaction: supervision horizon and conditioning schedule. Long-horizon supervision exposes content inconsistencies that remain locally plausible over short windows, while per-chunk conditioning allows instruction changes to be represented within the same rollout. These properties are then transferred to the few-step student through long-horizon distribution matching. 3 Evoke Evoke realizes long-horizon interactive generation through two complementary components: a bounded recurrent student with externalized world state, and a teacher designed for long-horizon interactive supervision. At each recurrent step, the three-step student generates one video chunk from a short local history, view-aligned geometry retrieved from an external camera-indexed world state bank, and a per-chunk text condition, without increasing the context length or positional span of any individual generation step as the session grows (Fig. 3). The teacher combines chunk-wise sparse attention with per-chunk conditioning to provide supervision across distant moments and changing instructions, and these capabilities are transferred to the student through a 30-second long-horizon distribution-matching objective applied under self-forced rollouts. The following sections formalize the recurrent formulation, motivate the required supervision horizon, and detail the teacher, world state bank, distillation procedure, and inference process. Figure 3: With bounded retention, per-step cost does not grow with session length. The camera pose reads the world state bank into the target view; the student injects that render at the coarsest of its three coarse-to-fine evaluations, alongside a short parametric history and the current text condition. The emitted chunk then updates the store and the history. 3.1 Recurrent Session Formulation 媒体内容 · 前往原文查看 Figure 4: An hour-long session stays bounded rather than degrading. One continuous 65.5 min Evoke rollout, three steps per chunk without classifier-free guidance (2619 chunks; faint per-chunk, bold two-minute mean). (a) Color is flat over the hour. (b) Scene identity plateaus at cosine 0.523, the level real footage scores against itself 60 s apart. (c) Opening-window shift stays far below unrelated-scene drift. A stability claim, not fidelity (n=1). Evoke represents long-running generation as a bounded recurrent process over video chunks. At recurrent step k, the student generates a chunk xk containing F=9 latent frames, corresponding to 36 pixel frames or 1.5 s at 24 fps. Each step is conditioned on a camera trajectory 𝒫k={Pt} (He et al. 2024; He et al. 2025a), containing camera-to-world extrinsics and intrinsics resampled to 24 fps, and a text condition ck that may vary across steps. A session x1:K is therefore formed by repeatedly applying the same fixed-shape generation interface for an arbitrary number of steps. The recurrent interface separates transient denoiser context from persistent geometric world state, the loop drawn in Fig. 3. Let hk denote a bounded local history and Mk the world state bank. At each step, geometry relevant to the current camera trajectory is first rendered from Mk, the next video chunk is then generated from the rendered geometry, local history, and text condition, and the resulting observation is written back to the store: rk=Read(Mk,𝒫k),xk∼pθ(⋅∣rk,hk,ck),Mk+1=Write(Mk,xk,𝒫k). (1) Both the local history and world state bank operate under fixed budgets. Consequently, extending a session increases only the number of recurrent calls, without increasing the context length, positional span, or computational footprint of an individual call. Figure 4 illustrates an hour-scale session under this formulation, and Sec. 4.2 evaluates the resulting runtime behavior. This bounded recurrence determines how Evoke can continue generating for long durations, but does not by itself ensure long-range content consistency or responsiveness to newly introduced instructions. These capabilities arise from two complementary sources: the supervision used to train pθ, which determines what long-horizon behavior the student can acquire, and the external state Mk, which preserves previously observed scene information across recurrent steps. The following sections analyze these two roles separately, beginning with the supervision horizon. 3.2 Long-Horizon Supervision The bounded recurrent formulation keeps each inference step independent of session duration, but it does not by itself prevent errors from accumulating across recurrent steps. The remaining question is how much of a rollout must be visible to the training objective in order to constrain long-range behavior. Let qθ denote the distribution of student trajectories x1:K and p the corresponding data distribution. Under self-forced distribution matching (Yin et al. 2024b; Yin et al. 2024a; Huang et al. 2025a), supervision is evaluated over windows of W consecutive chunks: ℒW(θ)=𝔼k[D(qθ(k:k+W−1)∥p(k:k+W−1))], (2) where D denotes the divergence whose gradient is estimated from the teacher and critic scores. By construction, this objective depends only on trajectory statistics observable within a W-chunk window. Two trajectories may therefore receive identical supervision at this horizon even if their behavior diverges over longer temporal spans. The supervision horizon W consequently determines not only how much generated history is exposed during training, but also which forms of rollout drift can produce a learning signal. This distinction is important because long-rollout errors arise at different temporal scales. Degradation drift, such as exposure shifts, saturation changes, and progressive texture degradation, eventually becomes visible within a local window. Its cause, however, may precede that window. During a self-forced rollout, the student conditions on its own generated history rather than ground-truth observations (Huang et al. 2025a; Cui et al. 2025), so earlier errors perturb the conditioning distribution encountered by subsequent denoising steps. A short window may capture the resulting degradation while exposing only a narrow range of the preceding perturbations that produced it. Increasing W therefore broadens the distribution of rollout-induced histories under which the student is supervised and improves its ability to remain stable under accumulated conditioning shift. Content drift imposes an additional constraint. Scene identity, object appearance, or spatial layout may evolve gradually while every short window remains individually plausible. Such errors become identifiable only when sufficiently distant moments are jointly covered by the supervision window (Cui et al. 2025; Liu et al. 2025). Long-horizon supervision thus serves two complementary roles: it enlarges the range of rollout perturbations under which the recurrent generator learns to remain stable, and it exposes long-range inconsistencies that cannot be identified from local trajectory statistics alone. Not all long-range inconsistency should be resolved by extending the supervision horizon. For temporal content drift, the correct evolution is governed by dependencies in the data distribution across distant moments; no particular past observation uniquely specifies the desired continuation. This failure is therefore naturally addressed through longer-horizon supervision. Spatial revisit inconsistency has a different source. When the camera returns to a previously observed surface, the relevant state has already been produced by the session, and the problem is to make that specific observation available again. Evoke accordingly assigns these two cases to different mechanisms: long-horizon teacher supervision constrains temporal evolution, while the world state bank preserves observed scene state and retrieves it through camera pose. A longer supervision window cannot substitute for an unavailable observation, and stored geometry alone cannot determine how unobserved content should evolve. Importantly, the required supervision horizon need not scale with the final session duration. The goal is not to train the student to recover from arbitrarily corrupted histories, but to make its recurrent dynamics stable over the range of perturbations encountered during normal generation. As the student becomes more robust to accumulated conditioning shift, its own rollouts remain within a progressively better behaved region of trajectory space, reducing the need to cover increasingly extreme failure states. Extending W should therefore provide large gains while it continues to expose new rollout perturbations and previously invisible long-range inconsistencies, followed by diminishing returns once the relevant temporal scale is covered. Evoke uses approximately 30 s of long-horizon supervision and tests this prediction by sweeping W: resistance to drift should improve with the supervision horizon and then approach saturation rather than scale proportionally with the desired rollout length. This analysis determines the requirements placed on the remaining components. The teacher must support sufficiently long windows at a cost compatible with repeated distillation queries, motivating the efficient long-sequence architecture in Sec. 3.3. Its supervision must also expose the time-varying conditioning encountered at deployment, since a teacher evaluated under a single global prompt cannot demonstrate how an instruction introduced mid-session should affect the rollout. Finally, previously observed scene geometry must remain explicitly accessible when revisited, which motivates the external world-state mechanism in Sec. 3.4. 3.3 Long-Horizon Interactive Teacher Figure 5: Chunk-wise sparse attention makes long, time-varying supervision affordable. Each query chunk of a 60 s clip attends to a fixed set of key sources and reads the text segment covering it, so attention cost grows linearly in length; the sampled timestep selects the high- or low-noise expert. Below: camera-control training and few-step distillation precede long-rollout distribution matching against the shared teacher/critic backbone. 媒体内容 · 前往原文查看 Figure 6: A student inherits photometric stability from a long-horizon teacher. Two three-step students with identical distillation recipes and teachers differing only in temporal horizon; clips, trajectory, prompt, seed and inference flags are shared, so only the checkpoint differs. (a) The short-horizon student settles at 74% of its opening brightness, the long-horizon one at 101%. (b) Seven of eight clips improve (Wilcoxon p=0.016). (c) Brightness separates the two most; sharpness not at all. Claim: exposure stability, not image quality. The previous section shows that the long-horizon behavior acquired by a few-step student is directly limited by the supervision expressed by its teacher. Evoke therefore treats the teacher not as a fixed source of high-quality scores, but as a model whose supervision is explicitly designed for long-horizon interactive generation. Figure 6 provides a controlled comparison: two students are distilled with identical recipes from teachers that differ only in temporal horizon, and the student supervised by the long-horizon teacher exhibits substantially greater stability over extended rollouts; Sec. 4.3 provides the full analysis. The Evoke Teacher is built on the 14B Wan2.2 A14B diffusion transformer (Peebles and Xie 2023; Wan et al. 2025a; Wan et al. 2025b) and retains its high-noise and low-noise experts selected according to the sampled timestep. The teacher and critic share the same backbone rather than residing as two separate models. Enabling a LoRA adapter (Hu et al. 2021) yields the critic, while disabling it yields the teacher, so the teacher and critic scores are always computed within the same expert. To make long-window supervision computationally practical, the Evoke Teacher partitions a sequence into chunks of nine latent frames and applies chunk-wise sparse attention (Jia et al. 2025; Cai et al. 2025). Each query chunk accesses a bounded set of sources: a first-frame global sink, local context with a one-frame overlap, spatially compressed nearby frames, a small set of selected distant frames, and a global state accumulated through linear attention (Yang et al. 2023; Team et al. 2025; Chen et al. 2025b), as illustrated in Fig. 5. Because the amount of information accessed by each chunk remains bounded, the resulting attention computation grows approximately linearly rather than quadratically with sequence length, as in other sparse-attention designs for long video generation (Li et al. 2025c), making repeated long-window scoring feasible during distillation. The same chunk partition provides an independent text-conditioning context for every chunk, allowing prompt changes to be represented within a single long sequence. This factorization also brings the teacher closer to the student’s chunk-wise recurrent inference, since local context and retrieved distant frames are drawn from the past while only limited non-causal paths remain. The mismatch between bidirectional teacher attention and causal student inference has been studied previously (Zhu et al. 2026); Evoke instead focuses on the complementary axes of supervision horizon and conditioning schedule. Evoke transfers this long-horizon supervision to the few-step student through full-window distribution matching distillation (DMD) (Yin et al. 2024b; Yin et al. 2024a). Training starts from one ground-truth prefix chunk x0gt followed by a self-forced rollout of 20 generated chunks x1,…,x20, yielding 21×9=189 latent frames, or 753 pixel frames corresponding to approximately 31.4 s. Since the student performs three function evaluations per chunk, constructing the rollout requires 60 student forward passes. Rather than sampling a short window from this trajectory, the teacher and critic jointly score all 189 latent frames, so the supervision horizon spans the complete self-forced rollout. Let x^0 denote the student prediction, sreal the teacher prediction, and sfake the critic prediction. We define the score difference Δ​s and normalize it by ν over the region Ω that receives the distribution-matching gradient: Δ​s=sfake−sreal,ν=meanΩ⁡[|x^0−sreal|],ℒgen=12​‖x^0−(x^0−Δ​sν)detach‖22. (3) The mask Ω excludes the ground-truth prefix and the first generated chunk x1. At this boundary, the teacher treats its first chunk according to the image-model first-frame distribution, whereas the student performs video continuation; directly matching the two introduces boundary flicker. The DMD gradient is therefore applied only to the remaining generated chunks x2,…,x20, and ν is computed over the same mask so that excluded frames do not affect the update scale. An important distinction is that the supervision horizon need not equal the gradient horizon. History is detached between consecutive rollout chunks, restricting each backward graph to a single chunk and allowing chunk-level gradients to be computed independently. This shortens the gradient horizon without shortening the supervision horizon: the teacher and critic still evaluate the complete 31.4 s trajectory jointly, so each local student update is derived from a distribution discrepancy defined over the full rollout. Long-horizon supervision therefore does not require backpropagation through the entire trajectory. Detachment bounds activation memory, while full-window scoring preserves the long-range training signal. The teacher must also express the time-varying conditioning encountered at deployment. At inference, the student models pθ​(xk∣x