acceptodds
Under review as a conference paper at ICLR 2027

Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation

Abstract

Interactive autoregressive video generation demands both low-latency rollouts and precise online control. Few-step distillation accelerates generation by reducing denoising steps, while online control imposes a causal constraint: frames and blocks should depend on history and controls available during generation. Existing video distribution matching distillation (DMD) pipelines, however, often supervise causal few-step students using bidirectional teachers that score complete clips. The score for a target can therefore depend on future frames and controls that were unavailable when the student generated it, misaligning teacher supervision with the student's causal information set. We introduce Context-Matched Distillation (CMD), a causal DMD framework that aligns teacher supervision with the information available when each target is generated. CMD replaces bidirectional full-clip scoring with a causal teacher that evaluates each target without access to future frames or controls. The same causal teacher initializes the few-step student, establishing a consistent causal formulation across teacher training, student distillation, and inference. Beyond aligning the temporal information boundary, Prefix Scoring matches supervision to the student's realized rollout context by evaluating each target under the cached student-generated prefix that produced it. Prefix Corruption further stabilizes training by perturbing unreliable prefixes produced early in training while preserving this target-context alignment. With a simple causal formulation, CMD naturally extends to frame-wise and chunk-wise generation, long video distillation, and camera-conditioned distillation. Experiments demonstrate state-of-the-art aggregate performance among autoregressive methods on both short- and long-video benchmarks, together with substantially improved adherence to time-varying camera controls.

Primary area:
generative models
Keywords:
Timestep Distillation, World Models, Video Generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_PJn7OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a method to achieve auto-regressive video generation by distilling a teacher video generation model. The key insight of the paper is the mismatch between teacher and student models, since the teacher sees the whole video, whereas the student only sees the past. The solution to this is to constrain the teacher to only see the past as well. Camera-conditioned causal distillation employs a similar approach, using only the camera embeddings available up to the current frame. The results are very impressive, with extremely smooth videos, and no visible temporal seams due to chunking, which may be the case in some methods.

Critical strengths

1/ I found the paper in general to be well-written and clear. This is appreciated, especially in this domain, which tends to contain a lot of technical parts and details. 2/ The central premise and solution of the method are clear, and make a lot of sense. This makes a strong argument in favour of the paper. 3/ The results in the supplementary material are visually very impressive, and I congratulate the authors on them. This is obviously a huge plus for the paper.

Critical weaknesses

1/ While the paper was in general well-written, I thought that the jargon was a bit difficult to follow when diving into the details. I understand that much of it is unavoidable, but I felt that some could be a bit clearer, and in general it bogged the reading/understanding of the paper significantly. Examples:

  • Line 180: "is rolled out on-policy". If I understand correctly, this is jargon borrowed from reinforcement learning, which the readers of the current paper are not necessarily familiar with, and which you have not, in any case, introduced. What is on-policy rolling out? I saw that it has indeed been used in other very recent preprints on autoregressive video generation, but I think it should be explained, otherwise the paper becomes difficult to read.
  • Throughout the paper, you use "scoring", for evaluating the score, ie the gradient of the log probability of the video. I do not think this is common jargon, I would advise changing it. For me, it makes things slightly cloudier than they have to be. Especially in this domain, I feel like too much new jargon is introduced, making the papers difficult to read. If I am wrong, and this is common in auto-regressive diffusion-based video generation, please point me to some (published) papers, and I will accept the jargon.
  • The use of the word "prefix". While I understand this is common terminology in LLMs, it is not so common in video generation models. While I understand you do not have space to define it in the abstract, you must define what it means in the beginning of section 3, otherwise some readers will have trouble following.

2/ I think you really need some notation for the different models, ie:

  • teacher
  • causal student variant
  • block causal student variant
  • prefix

In summary, I think that because the authors are so familiar with many different domains relevant to the present task, they have ommitted to define certain notions carefully. This, combined with the lack of notation in some cases, leads to long, technical sentences which are not easy to follow.

3/ Also, I think that the whole premise of the method, and the solution you propose, could have been explained a bit better. If I summarise it, it is: "during training of the student, the teacher sees the whole video, whereas the student is auto-regressive, and so only sees the past. This creates a mismatch between teacher and student. Our solution is to constrain the teacher to only see the past as well". (if this is incorrect, please correct me). I think a diagram could be of great help here. In particular, the paragraph "conditioning mismatch in causal training" is central to your premise, and slightly difficult to follow. Could you not put the diagram in the supplementary material illustrating which information is available for the teacher and the student ? And also maybe an illustration of your solution, CMD. It might help make your case clearer, and certainly more impactful.

4/ The beginning of Section 3 was difficult to follow for me. Some examples:

  • Line 182: "Building on base CMD". What is base CMD ?

  • Line 183: "the cached student-generated prefix used to generate it". By this, you mean the video generated by the student until time i-1, which as been put into the model to calculate K and V. Again, this could be clearer.

5/ I find it extremely odd that you put the related work in the supplementary material! This seems unfair: other authors will have made an effort to make space for the related work in the main paper, whereas you have bypassed it. My opinion would be to put it in the main paper.

Questions for authors

1/ In section 3.1, it seems like you present the conditioning mismatch as if this is the first paper to remark on it. Is this the case ? If not, can you cite the others (apart from Zhu 2026b, which you put). Rather than in section 3.1, this should be earlier, in related work, since it is your core premise.

2/ When looking at the factorisation in Equation 2, I felt that something was slightly odd. While I agree that the conditioning on B_{<i} is just a consequence of the definition of conditional probability, the conditioning on C1:NC_{1:N} turns into C≤iC_{\leq i}. This is not necessarily obvious. Indeed, you are saying that future controls are independent of generated blocks in B<iB_{<i}, which means that the user is not making future decisions based on what is happening in these previous blocks. This may be a hypothesis that helps simplify the model, but in practice is probably not true. I am not saying this should be changed, but the hypothesis should be made clearer. At first glance, it seems like what is being said is "the controls in the future are unknown, so therefore they are independent", but unknown does not mean independent. If I have misunderstood something, please explain further.

3/ Why do you not compare with other methods in your supplementary materials videos !? You have the videos, since you have the quantitative results, why not add a few (not all obivously) examples of other works ? This seems odd, since your results are so good.

Small comments:

3/ Line 139: you re-use xx to define it as a video (in equation 1 it was a generic datum). I think you could just remove xx here, since you never use it again. The confusion for me comes from the fact that everything which is a vector is in bold, and xx is normal font. I think it is a very good idea to put vectors in bold, but maybe you could just say x0:T\mathbf{x}_{0:T} without referring to xx at all. However, this is only a suggestion, and a very minor point, you do not have to take it into account.

4/ Line 147: Why do you use both "chunks" and "blocks", aren't they the same thing ? If so, I would only use one jargon (preferrably chunks), unless they are not the same.

5 Line 153: "We treat the input image I0I_0 as part of the ordinary causal history rather than something that is permanently retained beyond local attention windows." The meaning of this sentence is not clear to me.

6/ Equation 3. Is this equation useful ? Why is it useful to know that setting a block size of 1 just gives generation frame by frame ? This is obvious no ? You could equally say that setting K=TK=T gives a standard bidirection model, and it would not be that useful either. Maybe remove.

7/ Line 231: "These noised targets are arranged temporally", what does arranged temporally mean?

8/ Line 241: "Frozen causal teacher", is this the same one that you discuss in the beginning of the section 3.2 ? If so, it should be defined earlier.

9/ Equation 5: The notation C\mathcal{C} is not great, can be easily confused for the conditioning in Equations 2-3. I would change this.

10/ Line 315: "We unproject rays". This is not very clear. The whole subsection is not super clear for me. What is Et−1−1E_{t-1}^{-1} ? It is not precisely defined it seems. I feel like the authors have attempted to make the section formal, but not really gone far enough, leaving the explanation a bit muddy. I would suggest either re-writing it, or adding a fuller explanation in the supplementary.

2 comments
AuthorsOpenReview ↗

Response to reviewer PJn7

Thank you for the careful reading. We have uploaded a revised paper and supplementary material, with all changes marked in red. The revision adopts the suggestions.

1. Terminology and notation

"On-policy", "prefix", "scores a target" and "Base CMD" are now defined at first use at the start of Section 4. "Block" and "chunk" are declared interchangeable. The corruption operator is renamed so that it no longer collides with the controls CC. The frozen teacher is introduced where it is trained.

On "scoring": real and fake score networks are standard terminology in the DMD papers we build on (Yin et al., CVPR 2024; Yin et al., NeurIPS 2024). We agree the verb was undefined and now define it once.

Instead of a separate notation table, we define each term inline where it first appears, so that the definitions sit next to their use.

2. Premise and diagram

Your summary is correct. Figure 2 illustrates exactly this information-set difference, and the mismatch paragraph in Section 4.1 now points to it.

3. Related work in the supplementary

Agreed. The main paper now has a Related Work section (Section 2) that covers prior work on teacher-student mismatch (Causal Forcing, Causal Forcing++, Context Forcing); the extended survey remains in Supplementary S1.

4. Q2: the factorization in Eq. 2

No independence is assumed. The joint over frames and controls factorizes as ∏ip(Ci∣B<i,C<i)⋅p(Bi∣B<i,C≤i),\prod_{i} p(C_i \mid B_{\lt i}, C_{\lt i}) \cdot p(B_i \mid B_{\lt i}, C_{\le i}), and the model specifies only the second factor, so the user's controls are free to react to earlier frames. The only assumption is that BiB_i does not depend on future controls given the past. This strengthens the premise: when controls react to frames, future controls carry information about the present frame, which is exactly what a teacher given the full trajectory can exploit. We state this after Eq. 2.

5. Q3: videos of other methods

Added, as side-by-side comparisons with four baselines in the supplementary gallery, and as camera-controlled comparisons with SANA-WM.

6. Smaller comments

  • The frame-wise equation (former Eq. 3) is removed and the frame-wise case is stated inline.
  • The sentence about the input image is rewritten: it is an ordinary history frame that leaves the local attention window in long rollouts.
  • "Arranged temporally" is now "concatenated in temporal order".
  • In the camera section, the relative increment is described as the camera motion from frame t−1t-1 to frame tt, and the ray construction is spelled out.
  • Line 139: the reused xx is removed, and the video is written only as the bold latent sequence.

[Edit: in the latest revision the ablation tables are renumbered in the order they appear. Former Table 5 (short-video) is now Table 4, former Table 6 (long-video) is now Table 5, and former Table 4 (camera) is now Table 6. Point 5 now also mentions the camera-controlled comparisons.]

Ok, thank you for the changes made, I feel it is clearer now. In light of this increased clarity, I will raise my score to accept.

Reviewer_U88tOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Context-Matched Distillation, which addresses the context mismatching problem between the teacher and the student in the DMD process by adopting a causal teacher that matches the student's attention pattern instead of a bidirectional teacher. It further proposes Prefix Scoring and Prefix Corruption to further ensure the matching between the teacher and the student and mitigate the error accumulation problem. Experiments on I2V, long video, and camera-conditioned world models demonstrates the effectiveness of the proposed method.

Critical strengths
  1. The paper addresses a meaningful gap in causal video distillation: the mismatch between the bidirectional teacher and the causal student. Addressing such a gap can potentially benefit long-horizon rollout by mitigating error accumulation.

  2. The proposed method is validated on I2V, long video, and camera-conditioned video generation to validate its generalizability across different tasks.

Critical weaknesses
  1. Flaws in validation: There is no validation under the basic T2V setting to demonstrate the universality of the method. Many methods like Causal Forcing and LongLive are originally designed for T2V. The authors should add comparisons on T2V and explicitly state how they adapt these methods to I2V.

  2. Lacking video comparisons: The supplementary material does not provide video comparisons with baseline methods, making it hard to assess the improvement of the proposed method.

  3. Formatting concerns: The margins of section heads, figures, and tables are overly small, which may violate the formatting instructions. In addition, related work is not presented in the main text. There should be at least a brief related work section to ensure the integrity of the paper.

I may consider raising the rating if the authors properly address these concerns.

Questions for authors

In CausVid, it is claimed that the causal teacher's performance is inferior to that of the bidirectional teacher, which contradicts the paper's findings. More analysis and discussions on this problem is needed.

4 comments
AuthorsOpenReview ↗

Response to reviewer U88t

We thank the reviewer. We have uploaded a revised paper and supplementary material, with all changes marked in red.

1. T2V validation

Two effects are at play when a causal student is distilled from a bidirectional model. First, restricting attention to the past lowers generation quality, so a causal teacher is weaker than the bidirectional model it is fine-tuned from. CausVid and Self-Forcing keep the bidirectional teacher for this reason. Second, a bidirectional teacher scores each target using future frames and controls that were unavailable when the causal student produced it. This is the context mismatch our paper addresses.

Which teacher is better therefore depends on the conditioning. In image- and camera-conditioned generation, the input image already specifies the scene layout, so the causal teacher loses less quality. The context mismatch, however, is larger, because the bidirectional teacher also sees the future camera poses. This is the setting of interactive world models and the one our paper targets, and here the causal teacher gives the better student (Table 5).

In T2V, the first frame is generated from text alone. It has no history to match, and the causal teacher must decide the scene layout by itself, which is where the causal conversion costs the most quality. We evaluated this setting. A frame-wise T2V student distilled with CMD obtains a VBench Total of 81.84, compared with 83.01 for frame-wise Causal Forcing, 83.78 for Self-Forcing, and 84.34 for chunk-wise Causal Forcing, all re-evaluated under the same protocol. Relative to the frame-wise baseline, it improves subject consistency (92.8 vs. 90.7), background consistency (93.8 vs. 92.4), motion smoothness (98.5 vs. 96.7), and aesthetic quality (67.6 vs. 66.1). The remaining gap is in spatial relationship (65.8 vs. 71.4), scene (40.7 vs. 46.8), and imaging quality (66.9 vs. 70.1), which depend on layout and per-frame fidelity.

In T2V, these results suggest that causal teacher quality remains a bottleneck, making stronger causal T2V pre-training a promising direction beyond the scope of this work. Our paper demonstrates that a context-matched causal teacher produces a better student than a bidirectional teacher for image- and camera-conditioned generation. These are central capabilities of interactive world models, which must continue an observed scene and respond to controls online.

2. Adapting T2V baselines to I2V

The latent of the input image is set as a clean first latent frame, and the remaining frames are generated as usual. This is the I2V mode of the official Causal Forcing implementation (built on Wan2.1-T2V-1.3B). For chunk-wise baselines, the image latent is the clean first latent of the first chunk, and the model generates the rest of the chunk. No retraining is needed, because the first latent frame of the Wan VAE encodes a single image (as noted in Self-Forcing). Each baseline otherwise uses its released checkpoint and sampler. This is now stated in Section 5.1.

3. Video comparisons with baselines

The supplementary gallery now has a "Comparisons with baselines" section with side-by-side videos of our chunk-1 and chunk-4 models against Causal Forcing, Self-Forcing, LongLive and Context Forcing, on five scenes. Every method in a row receives the same image, prompt and seed.

4. Formatting

We have fixed the spacing around section headings, figures, and tables. A Related Work section is now in the main paper.

5. CausVid's finding on causal teachers

There is no contradiction: CausVid's finding concerns the first effect in point 1, the quality lost in causal conversion, and ours concerns the second, the context mismatch. We do not claim that the causal teacher is the stronger generator. Our finding is that, for supervising a causal student, a teacher that scores under the student's own context yields a better student than a stronger teacher that scores under a different one (Table 5: 88.33 vs. 82.65 Total).

CausVid's causal teacher is, like ours, a bidirectional model fine-tuned with a causal mask and per-chunk noise levels following Diffusion Forcing, so the teacher recipe is not the difference. The difference is the context under which a frame is generated versus the context under which it is scored.

In CausVid, the student predicts each chunk from noised dataset frames, while a causal teacher scores that same chunk given the student's predictions for the preceding chunks. The student therefore generates under a real history but is supervised under a generated one: the teacher is causal, but it is not context-matched. Using a causal teacher alone does not remove the mismatch, and the bidirectional teacher gives the better student there.

In CMD, the student generates each target from its own previous outputs, as in Self-Forcing, and the teacher scores that target under the same generated history. We added this discussion to Section 5.3.

Thank you for the detailed response. After considering the updated materials, I still have some further concerns:

  1. The paper's scope should be explicitly stated (I2V and optionally camera conditioning), and the discussion in item 1 should be included, at least in the appendix.

  2. The video comparisons still lack the camera-conditioned WM results with demonstration of GT camera trajectories, and prompts should be illustrated for I2V results.

  3. The spacing problem is still unattended.

AuthorsOpenReview ↗

Response to reviewer U88t

Thank you for the quick follow-up. We have uploaded a revision of the paper and the supplementary material that addresses the three points. Changes are again marked in red. Note that with the spacing restored, the ablation tables appear in a different order, and we have renumbered them accordingly: Table 4 (short-video), Table 5 (long-video) and Table 6 (camera).

1. Scope and the T2V discussion

The Introduction now states the scope explicitly: image-conditioned autoregressive generation, optionally with camera control. The discussion from item 1 of our response, including the T2V experiment and its VBench numbers, is now Supplementary Section S2.

2. Video comparisons

The supplementary gallery has a new section, "Camera-controlled comparisons". It shows six of our camera-control samples side by side with SANA-WM (limited by suppl. size), on scenes and trajectories of the SANA-WM benchmark. SANA-WM receives the same input image and the same camera trajectory, and the input trajectory is drawn next to the videos, with the current camera position and viewing direction marked as the video plays. SANA-WM is run with its released 60-step model and refiner; our models use 4 sampling steps.

The text prompt is now printed below every comparison video, for both the I2V and the camera-controlled comparisons.

3. Spacing

We apologize that this was not resolved in the first revision. All negative spacing has been removed from the paper: around the title, section headings, figures, tables, captions and equations. To stay within the page limit, the following text moved from the main paper to the supplementary, unchanged: three background passages of the Preliminaries (Section S3), and the long-video training details and corruption schedule (Section S4.1). No results, tables or figures were moved to the supplementary.

Thank you for the timely response. After checking the revised materials, I am happy to raise my rating to Weak Accept. Good luck!

Reviewer_cTBKOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper tackles a mismatch in how autoregressive video models get distilled. Usually a causal student generates frames one at a time, but it’s graded by a bidirectional teacher that can peek at future frames and future controls (like camera moves). So the student gets supervised with info it never had when generating. The authors fix this by making the teacher causal too, so it only sees what the student saw. They also have the teacher score each frame using the student’s actual generated history (Prefix Scoring), with a bit of noise added to that history so early messy rollouts don’t throw training off (Prefix Corruption). The same idea extends naturally to long videos and camera control, and they report state-of-the-art results among AR methods, especially lower camera-pose errors.

Critical strengths
  1. Clear and well-motivated problem. The paper points out a real mismatch, a causal student is supervised by a bidirectional teacher that can see future frames and controls. The issue is easy to understand and matters for interactive and camera-controlled generation.
  2. Simple and general method. The same causal model is used for teacher training, student initialization, and inference, so no extra initialization stage is needed. The same formulation extends to frame-wise and chunk-wise generation, long videos, and camera control with only small changes. Prefix Scoring is also implemented efficiently, scoring all targets in a single pass.
  3. Strong results. The method performs well on short-video and long-video benchmarks, with large gains over a bidirectional teacher in the ablations. It also improves camera control, the chunk-4 model has the lowest camera errors in Table 1, and the generated videos follow the camera inputs more closely. This is consistent with the motivation, since the teacher no longer sees future camera inputs.
  4. Clear writing and figures.
Critical weaknesses
  1. Related work and novelty are not clear. The main paper has no related-work section, and several baselines (Context Forcing[1], HY-WorldPlay[2], minWM[3]) are cited only in the supplement. Context Forcing also scores targets under the student’s generated history and injects accumulated errors into the teacher’s context, similar to Prefix Scoring and Prefix Corruption. Causal Forcing[4] also already points out the problem of bidirectional teachers. The authors should discuss them in the main paper and state what is new.
  2. The proposed components add little in the ablations. Most of the gain comes from switching to a causal teacher (gain more than 5 Total in Table 5). The paper’s own components help much less, Prefix Scoring alone lowers Total (88.33 to 87.25), and Prefix Corruption only beats Base CMD at one noise level (88.46 vs. 88.33). This makes it unclear how much Prefix Scoring and Prefix Corruption contribute.

[1] Chen, Shuo, et al. "Context forcing: Consistent autoregressive video generation with long context." arXiv preprint arXiv:2602.06028 (2026).

[2] Sun, Wenqiang, et al. "Worldplay: Towards long-term geometric consistency for real-time interactive world modeling." arXiv preprint arXiv:2512.14614 (2025).

[3] Zhao, Min, et al. "minwm: A full-stack open-source framework for real-time interactive video world models." arXiv preprint arXiv:2605.30263 (2026).

[4] Zhu, Hongzhou, et al. "Causal forcing: Autoregressive diffusion distillation done right for high-quality real-time interactive video generation." arXiv preprint arXiv:2602.02214 (2026).

Questions for authors

See weekness.

2 comments
AuthorsOpenReview ↗

Response to reviewer cTBK

We thank the reviewer for the careful assessment. We have uploaded a revised paper and supplementary material, with all changes marked in red.

1. Related work and novelty

We agree. The main paper now has a Related Work section (Section 2), and every baseline is cited in the main text (Section 5.1); the extended survey remains in Supplementary S1. On the two closest works:

Causal Forcing uses a causal teacher to initialize the student through ODE matching; the DMD stage that follows keeps a bidirectional teacher. The mismatch we address is in that DMD stage.

Context Forcing is the closest in spirit, and we say so in the paper now: it scores under student-generated history and makes the teacher tolerant of the student's errors. The primary difference is that it is a second stage applied to an already distilled student, with a separate long-context teacher trained for error tolerance. CMD uses one multi-step causal teacher to both initialize and score the student, scores every target under its own prefix in a single pass, and obtains robustness by corrupting the prefix at scoring time, with no teacher retraining. CMD also covers frame-wise generation and camera control, which Context Forcing does not address.

2. Contribution of the proposed components

We agree that most of the short-video gain comes from the causal teacher. That is our first contribution and the reason for the title. On short clips the entire context lies inside the scoring window, so Base CMD and Full CMD are expected to be close, and we do not claim a meaningful difference between them there.

Prefix Scoring matters when the context extends beyond the scoring window or carries controls:

  • Camera ablation (Table 6): adding the prefix raises Semantic from 0.10 to 0.24 and Total from 0.675 to 0.706.
  • Long-video ablation (Table 5): the prefix carries history across clip boundaries, which Base CMD cannot see, and raises Total from 69.37 to 70.02.

Prefix Scoring without corruption lowers the short-video score because early student prefixes are far from what the teacher saw during training. This is the case Prefix Corruption is designed for. We have made this explicit in Section 5.3.

[Edit: in the latest revision the ablation tables are renumbered in the order they appear. Former Table 5 (short-video) is now Table 4, former Table 6 (long-video) is now Table 5, and former Table 4 (camera) is now Table 6. The numbers in this response have been updated.]

Thank you for your detailed rebuttal. All of my concerns have been addressed, and the revised paper looks much better now. So I will raise my score to Weak Accept. Good luck!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.