acceptodds
Under review as a conference paper at ICLR 2027

A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models

Abstract

When a video model generates physically incorrect motion, did it fail to learn the correct motion, or did it learn it but fail to use it? We show the latter: the correct motion remains available inside the model and can still be made to control the generated video. We train on videos where red masses oscillate slowly and blue masses oscillate quickly, then test a red mass with fast observed motion. Even when the model generates slow motion in this conflicting case, a low-dimensional edit predicted from simple physical variables restores the correct fast motion. We call this ability causal writability. At fixed strength, we find a sharp depth boundary: the same edit changes the video before the boundary but not after it. This closure marks commitment for that write. The motion signal nevertheless remains, and a stronger downstream write can restore physical motion, while excessive gain overshoots. Early causal writability predicts which errors training later corrects: those errors are writable at more network depths than errors that persist. We reproduce both causal writability and its sharp closure in a pretrained 1.3B video model, supporting generality across model scale and training regime.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, video diffusion models, activation interventions, causal writability, shortcut learning, physical reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NwweOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work investigates causal writability in video generation models. Using toy spring, pendulum, and free-fall datasets, an analysis of token state edits reveals that causal writability disappears beyond a certain block. Experimental results further demonstrate that analyzing causal writability using only the top four principal components (PCA) is meaningful. However, even after token-edit writability is lost, causal writability can still be achieved through attention KV restoration.

Critical strengths

Spring, pendulum, and free-fall simulation datasets were constructed to investigate the causal writability of video Diffusion Transformers (DiTs) through various experiments. Beyond a toy DiT model, the pretrained Wan 1.3B model was adapted to the simulation datasets, demonstrating that causal writability can be achieved through single-head attention KV restoration.

Critical weaknesses

Fragmented Experimental Results and Lack of Logical Flow

  • Logical flow of the manuscript. Although Sec. 2–5 and Appendix B present various experimental results, the logical connections between the experiments and sections remain unclear. In particular, the manuscript does not establish a coherent progression explaining how each experimental finding motivates subsequent analyses. This fragmented presentation obscures the central claim of the paper.
    • Especially, although Sec. 3 introduces the first-harmonic map through a harmonic motion formulation, this analysis is not utilized in Sec. 4–5. Furthermore, the first-harmonic map has not been validated in experiments using the pretrained Wan model.
  • Reproducibility. The lack of detailed experimental setups and sufficiently informative figure descriptions raises concerns about reproducibility and makes it difficult to critically evaluate and reliably interpret the experimental results.
    • Further details are provided in the Questions For Authors.
  • Inconsistent Terminology
    • The inconsistent use of terminology throughout the manuscript makes it difficult to understand newly introduced concepts and interpret the experimental findings.
    • Terms such as first-harmonic map, target-specific phase, state model, frozen controller, low-order organization, small controller, boundary-state controller, executable controller, donor-free controller, fit-only phase controller, and state-conditioned Top-4 controller appear to refer to the same equation in Line 246.
    • Similarly, synthesized held-out writes, predicted write, predicted edit, phase-predicted edits, and fit-only state edit appear to describe the same controller-generated edit.
    • Likewise, full edit, full paired-difference edit, full paired difference, full write, and matched full edit appear to refer to the same non-PCA full edit.

Significance and Empirical Validation

  • Motivation. The central claim of this paper regarding the feasibility of physical correction is primarily motivated by the observation that "physics-follow rates span about 22 percentage points across the three training seeds" (line 258). However, this observation alone provides insufficient motivation for the feasibility of model correction through variations in training seeds. Furthermore, experiments based solely on Spring, Pendulum, and Free fall simulation data are insufficient to represent the broader scope of physical phenomena.
  • Logical gap and technical validity. The edit intervention (line 218) is used to determine causal editability/writability and closure. However, it remains unclear whether the observed effects reflect the limitations of simple token-state steering or an actual loss of physical information within the model. This ambiguity creates a logical gap in the argument leading to Fig. 6a, which motivates the need to restore lost physical information. Furthermore, although K/V restoration and head selection are introduced to mitigate this issue, experimental evidence and analysis supporting these design choices are lacking. The evaluation is limited to a comparison with simple MLP output restoration. In particular, no justification or analysis is provided for the selection of head 8.
Questions for authors
  • In Fig. 2b, what is the source population from which the "23/26 shortcut failures" were obtained, and how were these 26 cases selected?
  • In Line 248, how was the "direction-only model" constructed, and what are its inputs?
  • In Fig. 4a, how many valid rollouts were obtained for each checkpoint?
  • In Fig. 5a, where is the aligned performance reported, as mentioned in Line 358?
  • In Fig. 5b, which intervention path was used (Direct or Neutral-first), which checkpoint and intervention site were selected, and what is the sample size (n)?
  • Which version of the Wan 1.3B model was used? Specifically, was it Wan 2.1 (1.3B), and which checkpoint was employed?
11 comments
AuthorsOpenReview ↗

1. Research motivation (1/3): the question and the corrective edit

Thank you for the careful reading and for pointing out that the connections between experiments need to be clearer.

Our research examines the competition between motion information and color cues in cases where the model generates incorrect motion. We ask how the correct motion can still influence generation from inside the model, and how this influence changes with network computation and training.

We follow two lines: how correction changes with depth and learning, and why models with transferable corrections still generate differently.

A common starting point: predicting a corrective edit

We first tried activation patching. We held the observed motion and generation seed fixed and changed only color, obtaining paired runs that generated correct and incorrect motion. We then added the activation difference between the two runs to the incorrect run at the same layer and let generation continue.

The edit changes the internal states representing observed frames. At injection, the states for future frames of the incorrect run retain their current values. We can therefore test whether changing the information from observed frames, while retaining these future-frame states, allows subsequent computation to correct the generated motion.

After finding that the full activation difference corrects motion, we asked whether the activation changes responsible for correction have a structure with few dimensions shared across examples, and how this structure varies with physical state.

We found that retaining only the first few PCA coordinates of the activation difference can still restore correct motion. The activation changes needed for correction thus have a clear organization that can be described with few coordinates.

These coordinates also vary continuously with the boundary phase, computed from the last observed position and velocity together with the frequency of the input history. This leads to a further question: can this relation between edits and physical state directly predict how a new example should be edited?

We therefore fitted a small model on a set of paired runs, using sine and cosine terms to describe the relation and predict edit coordinates from the target motion and physical state. For a new trajectory, the predictor uses this phase and the requested motion family to construct an edit that corrects some errors without retraining the video model or providing that trajectory's matched correct run.

The PCA experiment tests how much correction is retained by a few coordinates. Predicting edits from physical state applies this structure to new trajectories. This is the motivation for the edit predictor.

New controls clarify what these edits use. We identified directions related to color from separate fitting pairs that differed only in input color, while their generated motion and appearance were nearly unchanged. Holding the corresponding coordinates at the values in the incorrect run reduced frequency recovery from 330/384 to 1/384. Edits constructed from different motion histories at the same input color recovered 236/384 cases under the criterion for nearly complete frequency recovery. We interpret the predicted edit using four PCA coordinates through learned associations between color and motion. Physical state predicts its coordinates. Details: supplementary-evidence/index.html, Q1.

AuthorsOpenReview ↗

2. Research motivation (2/3): writability across depth and learning

Line 1: writable depth and learning

We found that edits succeed at shallow layers and fail at deeper layers. In Spring, motion recovery drops sharply over a narrow range of network depth.

This leads to a measurable question: for a fixed editing method and strength, at which network positions can an edit still make the correct motion affect the output? We use causal writability to describe this ability and measure how it changes with depth.

We then found that the position of this transition also changes during training. This suggests studying the writable range as an internal feature that evolves with learning, and examining its relation to the model's motion predictions.

When comparing models at the same training stage, models with a larger writable range also tend to generate motion consistent with the observations more often. We then followed training over time to examine how overall motion prediction and the writable range of the remaining errors change together.

As training proceeds, the model generates correct motion more often overall, but the remaining errors become harder to correct with the same type of edit. This leads to a question about individual samples: since each example has its own writable range, are early errors with greater writability more likely to be corrected by later training?

We tracked these examples and found that errors corrected later are already writable at more layers early in training. Early writability can therefore predict which errors are more likely to return to the prediction consistent with the observed motion as training continues. The generated videos are still incorrect at that early stage, but their internal writable ranges already reflect differences in their later outcomes.

We also varied the length of the observed motion history and the order in which the model learns motion and its association with color. Both a longer motion history and learning motion before introducing the color bias keep the correct motion writable at deeper positions.

The Pendulum, Free Fall, and pretrained Wan [1] experiments further test whether edits predicted from physical variables and the loss of editing effectiveness with depth also occur in other dynamical systems and models.

References

[1] Team Wan et al. Wan: Open and Advanced Large-Scale Video Generative Models. 2025.

AuthorsOpenReview ↗

3. Research motivation (3/3): transfer across models and the attention mechanism

Line 2: transferable corrections across models

The second line starts from the observation that rates of following observed motion differ by about 22 percentage points across training seeds.

This leads us to ask whether models with very different generation behavior have learned different amounts of information about motion, or have similar useful structures that are used differently by later computation.

The analysis of edit coordinates gives us a concrete object to compare. At each model's writable layer, we fitted simple scaling and rotation transformations between their edit coordinates, then tested whether an edit from one model changes the motion actually generated by another.

We found that corrective edits transfer across models. Models with very different generation behavior can still have similar edit spaces that can be aligned and used to correct motion.

This leads us to examine how these similar edits affect generation in later computation, and why different models allow different outcomes to determine the output.

Where the two lines meet: attention writes from observed to future frames

The depth analysis leaves the question of why observed-frame editing stops correcting motion beyond a particular position. The analysis across training seeds leaves the question of why models with transferable correction structures still differ substantially in their unedited behavior.

Both led us to examine which computations write motion information from observed frames into future-frame states, and how this write affects the final motion.

We first tracked the changes introduced by a successful shallow edit. These changes remain measurable in future-frame states even at depths where a new edit to observed frames no longer corrects motion. We then examined how these changes affect the output through subsequent computation.

Starting from a run whose motion has been corrected by the edit, we first replaced the state for future frames after attention and the MLP output, separately, with their values from the incorrect run. We then replaced Q, K, and V for observed frames individually with their incorrect-run values, asking which replacements make the motion incorrect again.

These replacement tests identify the write through which a successful edit affects generation. We then asked the reverse question: in an incorrect run, can strengthening the write associated with the correct motion restore that motion?

We used the edit predictor established earlier to construct the change corresponding to the target motion to strengthen. In one model, strengthening the corresponding write in a single attention head restores correct motion for some failed examples, allowing the correct motion to regain influence over generation.

We further held a later edit fixed while changing earlier attention writes, recovering cases where neither intervention alone succeeds. This links later writability to the state for future frames formed by preceding computation.

Reviewer_2HVTOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper shows that incorrect motion in video generation does not necessarily mean the model has failed to learn the correct dynamics: relevant information may remain internally available without controlling the output. Low-dimensional activation edits can correct errors caused by color shortcuts, but their effectiveness drops sharply beyond a narrow “commitment boundary,” after which stronger interventions are required. This boundary depends on training and observation history, with the mechanism traced to self-attention pathways from observed to future frames and validated across multiple physical systems and an adapted pretrained model.

Critical strengths
  1. The main idea is interesting which is to separate “the information can be decoded” from “the information can still causally control generation.”
  2. Closure is a useful observation, where the alternative future does not disappear, but the same direct write stops working past a depth.
  3. The mechanism is linked to learning. Writable-layer count at 5K steps predicts which shortcut errors are corrected at 80K and 100K steps. Causal writability may be an early signal of whether training can still repair an error, not only a property of one forward pass.
Critical weaknesses
  1. The main experiments use controlled synthetic dynamics with a perfect color–speed correlation. Shortcuts in real video are messier, and a similarly sharp closure is not shown there.
  2. Pretrained validation uses one 1.3B architecture and a hand-adapted task. It does not show that natural shortcuts in large general video models share the same mechanism.
  3. The PCA directions and the low-dimensional controller may follow the joint color–motion manifold in the training data. Restoring frequency also pushes future color toward the training-associated color, so physics and appearance are not fully separated.
Questions for authors
  1. Build natural shortcuts in real or semi-real video, such as background versus speed, object category versus collision outcome, and road type versus vehicle motion, and test whether closure remains.
  2. Test more public pretrained video models, scales, and generation paradigms. Prefer errors the model produces on its own over a hand-adapted task.
  3. The figures in the paper are hard to understand, though the main idea of this paper is quite interesting.
3 comments
AuthorsOpenReview ↗

Motion editing in more realistic driving scenes

Thank you for encouraging us to test a more realistic setting and to clarify the scope of the results with the pretrained model. We ran an additional driving experiment and address the scene, model, and color–motion questions below.

Q: Can closure be tested in real or semi-real video with more realistic shortcuts?

A: At fixed observed motion, lighting can bias the predicted speed; shallow internal edits correct that bias and deeper edits lose their effect. We added this test in photorealistically rendered Virtual KITTI 2 [1] driving clips using Wan2.1-T2V-1.3B [2] fine tuned for the task.

Training pairs daylight with slow motion and sunset with fast motion. Evaluation pairs keep the vehicle, motion history, crop, and generation noise fixed while changing lighting. Training and evaluation use separate time intervals and vehicle events, with no shared source frames.

At 20K fine-tuning steps, switching daylight to sunset increased the measured progress rate by 0.25 to 0.28 across three seeds on jointly measurable validation pairs; separate test clips shifted in the same direction. This rate measures progress along the reference image path, with slow/fast targets of 1 and 2.

Paired observed-frame editing restores motion early and loses effectiveness later. All ten eligible validation model–input pairs, drawn from three vehicle events across three seeds, met the recovery criterion after block 1; none met it after block 19. All outputs at these two editing locations were measurable. Recovery declines more gradually than in Spring, but approaches zero at later editing locations.

We used the complete difference between activations for observed frames at fixed multiplier 1, leaving future-frame tokens unchanged at injection. Eligibility was determined from unedited outputs before the depth scan. The evidence page retains every eligible case, unmeasurable edited test outputs, selection rules, measurement coverage, and paired videos.

This adds a lighting cue and more complex scenes within the same model family adapted to the task.

Evidence: Driving comparisons, complete depth curves, settings, and videos (supplementary-evidence/r2/index.html).

References

[1] Cabon, Y., Murray, N., and Humenberger, M. Virtual KITTI 2. 2020.

[2] Team Wan et al. Wan: Open and Advanced Large-Scale Video Generative Models. 2025.

AuthorsOpenReview ↗

Controlled experiments and the choice of pretrained model

Q: Can you test more pretrained models and naturally occurring errors without task adaptation?

A: Thank you for raising this question. Our experiments use Wan2.1-T2V-1.3B [1] fine tuned for synthetic motion tasks and driving scenes. We tested how learning history affects predicted edits and writability across network depth.

After 10K steps on the same biased data, the model that learned motion first followed the observed motion on 80.5% of 128 fixed test inputs, versus 9.4% for direct adaptation. On a separate fixed set of 17 cases with slow motion as the target that remained wrong under both paths, learning motion first also allowed effective editing at later layers. The latter comparison keeps the failure identities fixed.

Controlled adaptation lets us hold motion history and generation noise fixed while changing the appearance cue. The video DiT setup lets us edit tokens for observed frames while keeping future-frame tokens unchanged at injection, and then measure the resulting motion in the decoded video. This makes comparisons across depths and learning histories interpretable.

The pretrained-model evidence is limited to task-adapted Wan; other model families and naturally occurring errors without task adaptation remain untested.

Q: The figures are hard to understand. How will their presentation be clarified?

A: We have standardized names across the text, captions, and figure labels. We use edit predictor, predicted edit, and full edit consistently, so the same operation can be followed across figures and experiments. We have also clarified the input conditions, intervention, sample sets, and measured outcome in the captions.

References

[1] Team Wan et al. Wan: Open and Advanced Large-Scale Video Generative Models. 2025.

AuthorsOpenReview ↗

Color and motion in the editing space

Q: Do the edits separate motion from appearance, or follow their learned association?

A: Thank you for raising this question. The new controls support influence in both directions between the cues: changes derived from either color or motion history can redirect the generated motion and appearance. We describe the predicted edit using four PCA coordinates in terms of the associations between color and motion learned during training.

The predicted edit using four PCA coordinates strongly depends on identified changes related to color. We identified these directions on separate fitting pairs by varying input color at fixed motion and noise, while generated motion and appearance remained nearly unchanged. Holding their coordinates at the values in the incorrect run reduced recovery from 330/384 to 1/384. Subtracting a projection onto a randomized basis, matched in Frobenius norm to the removed color projection, retained 328/384 recoveries.

Motion history also provides a useful edit at fixed input color. Red inputs with fast and slow observed motion both generate slow motion. We subtracted the slow history run’s activations from those of the fast history run, using only the observed frames. We amplified the difference and added it to the fast history run, keeping the identified color coordinates fixed. Across three models and both target directions, with strength selected on separate fitting cases, 236/384 cases met the criterion for nearly complete frequency recovery; 127 also passed an independent trajectory check that did not rely on red/blue tracking. The reversed difference and a randomized control with the same Frobenius norm gave no frequency recoveries in this test.

These internal interventions correspond to the competition between the cues in natural generation. With an ambiguous input color, the observed motion can prevail and pull the generated appearance toward the color associated with that motion during training. The edit derived from motion history also changes the generated motion and appearance. Together, the behavioral and internal experiments show how information from either cue can influence both outputs.

Frequency can also change while the red/blue category is retained. A separate Long history experiment uses an edit space with eight dimensions and a predictor fitted from frequency contrasts at the same input color to vary frequency within a band. Across 1,920 target requests on new inputs, the first edited outputs achieved mean absolute errors of 0.0305 rad/s in the fast band and 0.0211 rad/s in the slow band, while retaining the input’s red/blue category under the check of mean RGB.

We have revised the abstract, introduction, and Section 3 accordingly. Physical state predicts the edit coordinates. Detailed constructions and controls are reported in Appendices C.1 and C.2.

Evidence: Color dependence, same color history edits, natural behavior, and frequency control (supplementary-evidence/r2/index.html, Color and motion).

Reviewer_WM3vOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors construct a deliberately underspecified task: during training, red ⇔ slow (ω ∼ U[2.2, 3.0]) and blue ⇔ fast (ω ∼ U[5.2, 6.4]) are perfectly correlated. At test time they present a conflict input — red mass with fast observed motion. The model follows color and generates slow motion: a shortcut.

They then ask whether the correct motion was never learned or merely unused, via an activation edit. Only the 1088 condition tokens are modified ; the 1024 target tokens are untouched at injection, and the edit is applied once on each of 20 flow-matching calls.

Critical strengths

1.Technically elegant. 2. Separating decodable / naturally used / causally writable, with decoded pixels rather than a probe as the readout. 3.The most surprising and most "discovery-like" result; this alone earns my accept lean.

Critical weaknesses
  1. Is the edit writing physics, or writing color? The missing decisive control. The authors themselves report that in Spring the slow→fast edit "shifts future color toward blue," and in Free Fall that appearance moves toward the target color in all 64 outputs — "a physically parameterized edit to the coupled target future, not a color-disentangled gravity subspace." The most parsimonious reading is therefore: the edit flips the red cue to blue in representation space, and the model then applies its learned color→frequency rule. If so, the title should read "the shortcut cue is writable," not "the correct physics remains writable."

2."Commitment" does not survive, and contradicts Section 5. The abstract says the edit stops working past the boundary (closure marks commitment); Section 5 then shows that single-head V amplification at λ = 8 recovers 56.3%, and Appendix B.3 reports that access on fixed identities closes and reopens across nearby checkpoints (B13/B14). That is not commitment — it is a monotonically declining causal leverage / rising required gain with depth. The latter framing is more honest and more interesting (it can be plotted as a controllability-margin curve). I would ask the authors to downgrade "chosen future / commitment / point of no return" throughout to "declining causal leverage at fixed gain."

3 Controllability is binary mode switching, not physical control. The text concedes it does not establish reliable control of unsupported middle frequencies. But within-band continuous ω control (e.g. specifying a target inside [5.2, 6.4]) is in-distribution and entirely feasible; its absence is glaring. As it stands, the results are fully compatible with "the model has two mode-directions and the edit toggles between them," which is far weaker than "physical quantities are writable."

Questions for authors

see the above

3 comments
AuthorsOpenReview ↗

1. Does the edit recover motion information, or only change the color cue?

Thank you for the careful reading. Your questions about color, edit strength, and control of specific frequencies led us to run the additional tests below. We respond to the three points in separate comments.

Q1: Is the edit writing physics, or writing color?

A1: Thank you for raising this alternative explanation. The predicted edit using four PCA coordinates strongly depends on changes related to color, while motion history can also provide an effective edit at fixed input color. Together with behavior in natural generation, these results suggest that the learned association between color and motion can be used in both directions.

What the color control establishes. We identified directions related to color by varying input color at fixed motion and noise, using fitting pairs whose generated motion and appearance were nearly unchanged. Keeping these coordinates at their values in the incorrect run during the predicted edit using four PCA coordinates reduced recovery from 330/384 to 1/384. Subtracting a projection onto a randomized basis, matched in Frobenius norm to the removed color projection, retained 328/384 recoveries.

What motion history adds. We compared inputs with the same color but different observed motion. For example, fast and slow observed histories with red inputs both generated slow motion, so neither supplied a correct fast continuation. We subtracted the slow history run’s activations from those of the fast history run, using only the observed frames. We amplified the difference and added it to the fast history run, keeping the identified color coordinates fixed. Both target directions were tested, with strength chosen on separate fitting cases.

Across three models, 236/384 cases met the criterion for nearly complete frequency recovery; 127 also passed a stricter trajectory check using an independent frequency estimate and tracking without red/blue classification. The reversed difference and a randomized control with the same Frobenius norm gave no frequency recoveries in this test. The effect persisted when histories shared the same input color and matched final observed position and velocity; the separate run with the aligned color was used for selection and evaluation, not to construct the edit (Appendix C.1).

The internal and behavioral comparisons tell a consistent story. In natural generation, ambiguous purple can let the observed motion prevail and pull appearance toward the color associated with that motion during training. The edit constructed from histories with the same input color likewise often changes future appearance with motion. Information from either cue can therefore redirect the learned continuation.

We have revised the abstract and introduction around this interpretation: the predicted edit using four PCA coordinates relies strongly on changes related to color, its coordinates are predictable from physical state and the target, and differences in motion history can also supply a useful edit. The depth experiments ask when an intervention can still restore the motion indicated by the observed history.

Evidence: Q1 controls, tests with the same final observed position and velocity, behavior, and videos (supplementary-evidence/index.html, Q1).

AuthorsOpenReview ↗

2. Why does editing become less effective at later layers?

Q2: Does “commitment” hold when stronger interventions can recover motion?

A2: Thank you for pointing this out. We now use “writability closure” in place of “commitment” and “point of no return”. We define closure as the loss of motion recovery with depth under a specified intervention and gain, at a given checkpoint. Section 5 tests a different intervention by amplifying V in a single attention head. Further training can also shift the boundary or make an edit effective again at a previously ineffective layer.

Increasing gain helps some cases near closure, but does not generally remove the depth dependence. We scanned the same intervention on the observed states across 31 network sites and gains from 0 to 32. Higher gains recover some failures near the boundary, while excessive gains can reduce recovery. Under the criterion for nearly complete frequency recovery, no tested gain succeeds for fast targets at B12 or deeper in any of the three models; some slow targets recover later. Appendix C.3 reports the full curves of recovery by depth and gain and the minimum successful gain on the tested grid.

We then asked whether the states for future frames formed by earlier layers affect the same later edit. An edit to observed frames leaves the current states for future frames unchanged at injection. Subsequent computation must therefore use the edited observations to revise states that earlier layers have already formed.

To test this, we fixed the later edit to observed frames after B9 at gain 1.5 and changed only the earlier attention updates to future tokens in B6 to B9. We recorded K/V for the observed frames from a separate run in which their states had been replaced at a shallow layer. At each earlier block, we computed the attention output for future tokens twice: once normally, and once using these reference K/V for observed frames while keeping the current Q/K/V for future frames unchanged. We subtracted the first output from the second and added this difference to the normal update to future tokens. This procedure was applied on all 20 sampling calls.

At B9, both conditions impose exactly the same activations for observed frames; the states for future frames retain the effects of their different preceding computations. We then compare the generated motion.

For Model B’s 16 cases with fast motion as the target, recovery to the target frequency band was 1/16 with the later edit, 3/16 with the earlier intervention, and 13/16 with both. When only the earlier intervention is applied, B9 resets the state for observed frames to the incorrect reference (g=0). A random earlier change matched to the added update’s Frobenius norm over future tokens and channels, combined with the later edit, recovered only 1/16. All combined recoveries with fast motion as the target passed the criterion for motion quality in Appendix A. Ten of the 16 cases failed under either intervention separately but succeeded with both; the corresponding counts were 10 and 7 in Models A and C.

Thus, the same later edit to observed frames can recover motion after earlier attention updates are changed. These results extend the strength analysis: later correction depends on both the intervention gain and the state for future frames formed by preceding computation.

Evidence: Curves showing recovery across depth and gain, plus tests with the later edit held fixed (Appendix C.3) (supplementary-evidence/index.html, Q2).

AuthorsOpenReview ↗

3. Can the edits control specific frequencies within a training band?

Q3: Can edits control specific frequencies within a training band, rather than only switch between fast and slow?

A3: Thank you for suggesting this test. Yes. The predicted edit using four PCA coordinates can reach multiple distinct frequencies within a training band after gain calibration for each input.

In Short, we varied only the gain of the predicted edit using four PCA coordinates from Section 3. The space, predictor, and editing sites were unchanged. For each fixed input and generation seed, up to 16 calibration videos measured the relation between gain and frequency before evaluating the requested numerical targets. We then selected a gain and generated a new video for each target, without a target reference run or further feedback for each target.

The same input can reach several distinct frequencies within the same band. In Models A/B/C, respectively, 85/128, 60/128, and 58/128 inputs reached at least three of five fixed targets. Their acceptance intervals do not overlap, so these successes require distinct generated frequencies while the input, generation seed, and edit direction remain fixed. This directly demonstrates control beyond selecting the fast or slow category.

Across five fixed and three random targets per input, 1,652/3,072 requests (53.8%) succeeded, including unsuccessful and invalid outputs in the denominator. The tested bands were [2.2, 3.0] and [5.2, 6.4] rad/s. Success required meeting the validity and motion quality criteria in Appendix A, membership in the target band, and absolute frequency error below 0.08 rad/s for slow or 0.12 rad/s for fast targets.

In Long, we predicted numerical edits from the observed video and requested frequency. With 65 observed frames, we fitted an edit space with eight dimensions and a predictor from frequency contrasts at the same input color, using separate fitting and validation data. Given the observed video and requested frequency, it produces an edit without feedback from outputs during testing. Across 1,920 requests on new inputs, the first edited outputs achieved mean absolute errors of 0.0305 rad/s (fast) and 0.0211 rad/s (slow).

Evidence: Data for requested frequencies, calibration and videos (supplementary-evidence/index.html, Q3).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions and additional experiments

We thank the reviewers for reading the paper carefully and for the helpful suggestions. The discussion led us to new controls and a clearer account of how the experiments fit together. Our work asks when observed motion can still influence a generated future through internal editing, and how this ability changes with network depth and learning history. The four updates below clarify that question and strengthen its evidence; responses to individual reviewers provide the details.

1. Closure is specific to an intervention and its strength.

A write can lose its effect with depth while another intervention still restores motion. We have therefore removed unrestricted claims of irreversible “commitment” or a “point of no return,” and now describe closure for the specified edit and gain. We have also standardized the names, clarified the connections between experiments, and specified the relevant sample sets and intervention settings.

2. Color and motion history can both redirect generation; frequency editing goes beyond a mode switch.

New controls show strong dependence on changes related to color in the predicted edit using four PCA coordinates. Complementary edits constructed from different motion histories at the same input color can also correct motion, even when neither reference run generates the desired continuation. Together with the behavioral results, this supports use of the learned association between color and motion in both directions.

We also tested specific frequencies within each training band. Calibrating the gain of the predicted edit using four PCA coordinates lets the same Short input reach multiple distinct frequencies within a training band. In Long, a predictor of eight edit coordinates fitted from frequency contrasts at the same input color produces numerical edits for new test inputs without feedback from outputs during testing.

3. The same later edit can become effective after earlier attention writes change.

Before a later edit, attention has already used the observed frames to update future-frame states. We held the later edit to observed frames fixed and changed these earlier updates. Their combination recovered cases where either intervention alone failed, whereas random changes matched to the earlier updates in Frobenius norm did not improve recovery. These results show that later correction depends on the state for future frames formed by preceding computation. The scan across depth and gain also shows that increasing gain helps some cases near closure but does not eliminate the tested depth dependence.

4. Motion editing also loses effectiveness with depth in driving scenes.

In Wan2.1-T2V-1.3B [1] fine tuned for the task on photorealistically rendered Virtual KITTI 2 [2] clips, changing lighting at fixed motion history biases the generated motion. We also observed causal writability and its closure with depth in these driving scenes. This adds a different appearance cue and more complex scenes within the same model family, while preserving controlled pairing.

These additions explain what the edits use, what they control, and why their influence changes during computation. They complement the finding that early causal writability predicts which shortcut errors later training corrects.

Blue text in the revised manuscript marks added experimental paragraphs and their accompanying explanations.

Supplementary evidence: extract the accompanying ZIP and open supplementary-evidence/START_HERE.html. Motion editing and depth controls (supplementary-evidence/index.html) · Driving experiment (supplementary-evidence/driving/index.html).

References

[1] Team Wan et al. Wan: Open and Advanced Large-Scale Video Generative Models. 2025.

[2] Cabon, Y., Murray, N., and Humenberger, M. Virtual KITTI 2. 2020.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.