- 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
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.
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.
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.