acceptodds
Under review as a conference paper at ICLR 2027

MotionSCM: Structural Causal Modeling for Interventional Human Motion Generation

Abstract

Existing human motion control interfaces specify semantic or kinematic properties without exposing cross-joint coordination as an independently editable mechanism. This makes it difficult to isolate an action's dependence on a particular coordination rule and trace how changes to that rule propagate across the body. We introduce , a motion generator for studying these dependencies through mechanism interventions. Its time-varying structural causal model uses explicit structural assignments to generate each motion channel. Under stated causal assumptions, post-training edits replace selected assignments while holding generator inputs and unedited mechanisms fixed. Per-channel gain constraints preserve rollout stability within the admissible operator set. On professional MoCap data from six industrial and craft processes, the learned structure varies across actions and individual motions. Interventions on the same rule produce distinct action-dependent cross-joint responses, with lagged recurrence sustaining the transmitted changes. MotionSCM retains competitive generation quality on three action-conditioned benchmarks. Structural ablations further show that improved generation fit need not preserve intervention behaviour. These results establish coordination as an explicit object of causal analysis within the motion generator. Code is available at https://anonymous.4open.science/r/MotionSCM-736C/.

Primary area:
causal reasoning
Keywords:
Human Motion Generation, Structural Causal Models, Controllable Generation, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NMW2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

MotionSCM introduces a motion generator with explicit, editable coordination rules between joints. A latent-conditioned structural decoder supports mechanism interventions, while gain constraints ensure bounded rollouts and intervention effects under stated assumptions.

Critical strengths

The explicit coordination rules provide an interesting way to control and study generated motion. The paper supports the design with theoretical guarantees, propagation controls, and structural ablations, while maintaining competitive generation quality across three datasets. The scope of its causal interpretation is clearly stated.

Critical weaknesses

The main cross-joint response is measured only over frames 6–25 of a 180-frame sequence (Sec. 4.2), and its magnitude depends on low-ROM filtering (Table A3), although action rankings remain correlated. Full-sequence results would help clarify how these effects develop over time. I consider this a limitation of the analysis rather than a major concern about the method.

Questions for authors

Why is the cross-joint response measured only over frames 6–25? Do similar response patterns persist over the full motion sequence?

3 comments
AuthorsOpenReview ↗

Response to Reviewer NMW2: Why frames 6–25, and does the response persist over the full sequence?

Thank you for your careful and encouraging review, and for your positive assessment of the method and its theoretical and experimental support. In short, we measure frames 6–25 because this window of four lag cycles covers the transition into the edited dynamics (Proposition 1). The edit's effect does persist to frame 180, as a constant shift in the other joints' mean.

Why frames 6–25. Appendix A.2.2 (paragraph "Intervention metrics", p. 20) defines this window, WrW_r, as the first 4k=204k = 20 solved frames after the k=5k = 5 seed frames. The four-cycle length follows Proposition 1 (Sec. 3.2). Under the same, possibly edited, solved operators and drive, the influence of the initial history is multiplied by at most gˉ\bar g every kk frames, so at most gˉ4≈0.1\bar g^4 \approx 0.1 remains after four cycles. Empirically, on the same 2,400 pairs as Table A3, the RMS distance of the edited-minus-base difference from its steady level (its mean over frames 26–180) reaches the range it keeps over frames 26–180 by frame 25.

Frame61015202526–180 (per-frame range)
RMS distance to the steady level0.00360.00200.00080.00040.00030.0002–0.0004

Over the full sequence. On the same pairs, we report mean ± SD over the 3 training seeds.

FramesRelative ROM change (Sec. 4.2)Trajectory RMS, edited minus base (normalized channels)
6–250.0482 ± 0.01040.00269 ± 0.00068
6–1800.0103 ± 0.00220.00322 ± 0.00081
26–1800.00281 ± 0.000590.00328 ± 0.00083

After frame 25, the edited-minus-base difference is almost entirely a constant shift in the other channels' mean, on top of their ongoing motion. Range of motion (max⁡−min⁡\max - \min) is unchanged by a constant shift, so the late relative ROM change is small, while trajectory RMS, which we use alongside ROM (Appendix A.2.2), captures it. The per-action results of Sec. 4.2 (Fig. 5) are measured in frames 6–25, where the range of motion changes as the effect builds up.

Low-ROM filtering. As you note, excluding low base-ROM entries reduces the magnitude of relative ROM, while action rankings remain correlated (Table A3). The relative ROM column below restates Table A3. We add the trajectory RMS, which has no per-channel denominator, to check whether the size of the edit's effect also depends on these exclusions (lowest 10% or 25% of base-ROM entries, frames 6–25, mean ± SD over the 3 training seeds).

Excluded base-ROM entriesRelative ROM changeTrajectory RMS, edited minus base
None0.0482 ± 0.01040.00269 ± 0.00068
Lowest 10%0.0286 ± 0.0066 (−41%)0.00244 ± 0.00065 (−9%)
Lowest 25%0.0186 ± 0.0046 (−61%)0.00215 ± 0.00060 (−20%)

The trajectory RMS changes far less, so the size of the edit's effect does not hinge on the low-ROM entries.

We will add these tables to the appendix of the revised version, and we would be glad to discuss any remaining point. Thank you again for your valuable time and consideration.

Thank you for the clarification and additional analyses. I will maintain my rating of 3.

Thank you again for your encouraging review and the helpful suggestion. We have uploaded a revised version with all changes marked in blue. We sincerely appreciate your time and kind feedback.

Reviewer_hDF9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MotionSCM introduces a latent-conditioned structural decoder for human motion generation, with explicit contemporaneous and lagged dependencies between motion channels. Selected dependencies can be edited while keeping the latent inputs and other mechanisms fixed. The authors derive sufficient conditions for bounded rollouts and evaluate generation quality, intervention responses, and editing locality on AIMove, HumanAct12, and CMU MoCap.

Critical strengths
  1. An explicit intervention interface with stability guarantees. The decoder exposes channel-level assignments that can be modified independently, with sufficient gain constraints for bounded rollouts. This provides a concrete and interpretable alternative to editing latent features or attention weights.

  2. Useful empirical analysis beyond generation quality. The paper includes multiple training seeds, propagation ablations, and structural comparisons. MotionSCM also achieves the lowest FID among the evaluated generation baselines on all three datasets, showing that the structural restrictions remain compatible with competitive generation.

Critical weaknesses
  1. The learned mechanisms are insufficiently validated as reproducible descriptions of coordination. The paper correctly limits its causal interpretation to the generator, but reconstruction quality and internally consistent interventions do not establish that the decomposition into coupling operators and latent-driven inputs is uniquely or reliably determined. The reported action-response ranking correlations across training seeds decrease from 0.63 at (k=5) to 0.37 at (k=14), despite improved FID. This matters because action-specific coordination analysis is a central contribution. A controlled benchmark with known mechanisms, or stronger evidence of robustness to alternative decompositions and ordering choices, would make this contribution more convincing.

  2. The editing comparison does not establish an advantage in meaningful motion control. MotionSCM uses a HumanAct12 checkpoint, while MotionCLR and SALAD use HumanML3D checkpoints, with only five source motions per method and action. Matching arm displacement helps, but differences in source motions and training data remain. Moreover, lower off-target response establishes locality, not whether the resulting coordination is appropriate. Appendix Table A6 also reports higher bone-length variability for MotionSCM than MotionCLR across all three actions. Stronger evaluation of edit correctness and motion plausibility is needed to establish the practical benefit of the structural interface.

Questions for authors
  1. What constrains the allocation of motion information between the latent-conditioned drive and coupling operators? Can similarly fitting models produce substantially different intervention responses?

  2. How sensitive are the learned coordination patterns and action-response rankings to alternative contemporaneous orderings?

  3. Can the editing advantage be demonstrated under matched training data or comparable source motions, while evaluating motion plausibility and intended edit success alongside locality?

6 comments
AuthorsOpenReview ↗

Overview of our responses (1/5)

Thank you for the insightful review and for the specific tests you propose. We are glad you see our central contribution, an explicit intervention interface with stability guarantees, as a concrete alternative to editing latent features or attention weights.

We agree that reconstruction quality does not make the decomposition unique, and that editing should be compared on matched data and sources with success and plausibility alongside locality. We answer the following questions with retrains under the paper's configuration and two alternative orderings, and with such an editing comparison.

  • Are the learned mechanisms reproducible descriptions of coordination (W1)? An independent three-seed retrain of the paper's configuration reproduces the three-seed-mean ranking of the Sec. 4.2 per-action response, so this ranking is not specific to the paper's three trained models.
  • What constrains the allocation of motion information between the latent-conditioned drive and coupling operators, and can similarly fitting models produce substantially different intervention responses (Q1)? Three structural constraints defined in the paper bound how motion is allocated between the drive and the coupling. Single models with similar fit can still respond differently.
  • How sensitive are the learned coordination patterns and action-response rankings to alternative contemporaneous orderings (Q2)? A standard feedback-arc-set ordering computed from the same learned coefficients gives substantially different orders, yet its per-action rankings agree with those under our rule about as closely as seeds within a rule agree, whereas one random permutation lowers both coordination-pattern readings and changes the per-action rankings beyond seed variation.
  • Can the editing advantage be demonstrated under matched training data or comparable source motions, while evaluating motion plausibility and intended edit success alongside locality (W2, Q3)? Yes. With identical training data, checkpoints, and source motions, our mechanism edit nearly always reaches its target and stays inside plausibility tolerances. Its off-target-to-target RMS ratio is over an order of magnitude lower (paired median per seed) than under latent editing at the same target.
AuthorsOpenReview ↗

Response to W1 (2/5): Reproducibility of the per-action responses

Retraining and ordering. An independent three-seed retrain of the paper's configuration reproduces the three-seed-mean ranking of the Sec. 4.2 per-action response at Spearman 0.755 [0.482, 0.880] over the 25 actions, so this ranking is not specific to the paper's three trained models. Under a standard feedback-arc-set ordering computed from the same learned coefficients, single-model rankings agree with ours at 0.555 [0.334, 0.700], close to the average agreement between seeds of one rule, and show no detected change (Q2).

Known-generator benchmark. For the controlled benchmark that W1 suggests, we used the first of the paper's three models as a generator of our model class whose operators are known exactly. We replaced every training and validation clip by its rollout with the same action and length and trained three models with the paper's configuration on this motion alone, so no intervention enters their training data or model selection. We compare per-action rankings under the right-forearm row attenuation of Sec. 4.2 and take as reference how closely each side agrees with itself, through the generator's retest on fresh prior draws and the three models' agreement with one another.

Per-action rankings, 25 actionsSpearman [95% CI]
Generator with itself, on fresh prior draws0.947 [0.831, 0.985]
Three models with the generator (mean over models)0.513 [0.265, 0.681]
Three models with one another (mean over pairs)0.408 [0.164, 0.603]
DD, second row minus the mean of the first and third−0.165 [−0.368, −0.024]

The retest shows that evaluation draws barely change the ranking. The 95% interval of DD lies below 0, so single-model rankings differ from the generator's. On real motion, too, single-run rankings agree only in part across training seeds (our reply to Q1).

Horizon. The horizon result cited in W1 is the reason we use k=5k = 5. As reported in Sec. 4.4 and Appendix A.3.2, longer lag windows lower FID without making action-specific responses more reproducible.

AuthorsOpenReview ↗

Response to Q1 (3/5): What constrains the allocation, and can similarly fitting models respond differently?

Q1 asks what constrains the allocation and whether similarly fitting models can respond differently. We address both parts with an independent three-seed retrain of the paper's configuration. For the first part, we test whether the coupling allocation learned under the paper's structural constraints repeats in the retrain, using the share of lagged weight that stays within-channel in each retrained model, as reported for the paper's models in Sec. 4.1. For the second part, we compare, within each set of three training seeds, the action-averaged response of Sec. 4.2 at its fixed strength, for the paper's three models and for the retrain.

Three training seeds eachLagged weight within-channelAction-averaged response, largest over smallest seed
Paper's models94% (Sec. 4.1)1.7-fold
Independent retrain93.7–94.3%1.8-fold

The retrain reproduces the within-channel share, so it is not specific to the paper's three trained models. The answer to the second part is yes. With overlapping FID intervals across training seeds, the action-averaged response still differs by the factors in the last column, so similarly fitting models of the same configuration can differ in how strongly they respond to the same intervention.

Three structural constraints defined in the paper bound how motion is allocated between the drive and the coupling.

  • Drive rank. The drive is confined to an RR-dimensional subspace of the 81 channels (Sec. 3.1; R=32R = 32, Appendix A.2.3).
  • Gain caps. The L1L_1 row caps of the stability guarantee (Sec. 3.2) bound what the coupling carries, and the learned operators use nearly all of this budget (total coefficient mass 52.1 of 53.5, Appendix A.3.2). Without the gain constraints, the lagged cross-joint share of the paper's configuration rises from 5.5% to 70.6%, as reported in Sec. 4.4.
  • Acyclic ordering. It limits which channels can drive which within a frame.

As reported in Appendix A.3.2, single-run rankings agree only in part. For this reason, each per-action response of Sec. 4.2 is a three-seed mean with its seed range (Fig. 5).

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

MotionSCM is an action-conditioned motion generator with explicit, time-varying dependencies between motion channels. A latent flow prior supplies inputs to a structural decoder, which generates motion recursively through contemporaneous and lagged couplings. Selected equations can be modified while keeping the latent inputs, drive, and other equations fixed. An acyclic ordering defines the contemporaneous dependencies, while gain constraints provide stability guarantees for admissible edits. Experiments on AIMove, HumanAct12, and CMU MoCap evaluate generation quality, intervention responses, and the effects of structural constraints.

Critical strengths
  1. The intervention interface is explicit and useful. The method exposes channel-level equations and allows selected mechanisms to be changed while preserving the remaining assignments. Combining this interface with a conditional generative prior and recursive propagation makes it possible to study how local edits affect the generated motion.
  2. The stability analysis is supported by substantial experiments. The gain constraints provide clear sufficient conditions for bounded rollouts under admissible edits. Competitive generation results, three-seed evaluations, and fixed-target propagation ablations support the implementation. The authors also appropriately limit their causal interpretation to the learned generator.
Critical weaknesses
  1. The empirical benefit over GOM-based coordination models is not sufficiently established. GOM already provides explicit, time-varying coordination models and motion forecasting, and subsequent work supports coefficient editing. MotionSCM adds a conditional generative pipeline, explicit intervention semantics, contemporaneous DAG structure, and stability constraints. These are meaningful distinctions, but the current comparisons do not isolate their practical benefit over a closer coordination-based model. The direct-decoder ablation removes the structural intervention interface entirely, while the attention-based editors address a different form of control.
  2. The robustness of the coordination interpretations remains only partly established. I find the generator-level intervention semantics coherent and appreciate the seed, horizon, and budget analyses. However, Appendix A.3.2 reports lower reproducibility of action-response rankings at longer horizons, despite improved FID. Since the coupling operators and drive are jointly learned from latent inputs, it remains unclear how consistently similarly well-fitting configurations recover these coordination findings. This limits the explanatory value of the results without invalidating the model’s internal SCM semantics.
Questions for authors
  1. Comparison with a closer structural baseline: Could you provide a controlled comparison with a GOM-inspired lag-only decoder, for example by retraining with A0=0A_0=0, while retaining the conditional latent prior, drive architecture, and recursive rollout, with comparable capacity and stability bounds? Reporting generation quality and intervention behavior at matched target effects would help establish what the contemporaneous mechanisms add. This would be a controlled architectural comparison rather than a reproduction of the original GOM.
  2. Sensitivity to the drive: Appendix A.2.3 fixes the drive rank at R=32R=32. Beyond the reported horizon and budget studies, how do generation quality and per-action intervention-response rankings change when RR is varied? This would help assess whether the coordination findings depend strongly on the capacity assigned to the latent-conditioned drive.
  3. Sensitivity to the ordering rule: Appendix A.3.2 acknowledges that removing DAG orientation also changes the final support density. Could you retrain with an alternative acyclic-ordering rule at matched edge density and the same gain budgets, and report generation quality and agreement in per-action response rankings? This would separate sensitivity to the ordering heuristic from the density change in the existing ablation.
5 comments
AuthorsOpenReview ↗

Overview of our responses (1/4)

Thank you for the insightful review and for specifying the controls so precisely. We are glad you value our central contribution, explicit cross-joint coordination rules that can be edited one at a time within a provable stability bound.

We agree that a GOM-inspired lag-only decoder is the right control for what the contemporaneous mechanisms add, and that fit alone does not show that the coordination findings are reproducible. We answer the following questions with three controls, for which we trained 24 new models in eight configurations.

  • What do the contemporaneous mechanisms add over a GOM-inspired lag-only decoder (W1, Q1)? They add an explicit, editable rule for direct cross-joint coupling within the stability bound. At matched target effect and with no resolvable difference in FID, actions in the paper's configuration depend several times more on the edited forearm rule than in a lag-only decoder (A0=0A_0 = 0) with matched capacity and gain ceilings, whose lagged operators do not take over the coupling.
  • How consistently are these coordination findings recovered, and how do generation quality and per-action intervention-response rankings change when RR is varied (W2, Q2)? An independent three-seed retrain of the paper's configuration reproduces the three-seed-mean ranking of the Sec. 4.2 per-action response. With the drive rank halved to 16 or doubled to 64, all FID intervals overlap, and the per-action rankings agree with those of the paper's R=32R = 32 models about as closely as seeds agree with one another.
  • Can sensitivity to the ordering heuristic be separated from the density change in the existing ablation (Q3)? At matched edge density and gain budgets, a standard feedback-arc-set ordering computed from the same learned coefficients gives substantially different orders, yet its per-action rankings agree with those under our rule about as closely as seeds within a rule agree, whereas one random permutation changes the per-action rankings beyond seed variation.
AuthorsOpenReview ↗

Response to W1 and Q1 (2/4): Benefit over GOM-based coordination models

Lag-only control. To isolate what the contemporaneous mechanisms add, we retrained a GOM-inspired lag-only decoder (A0=0A_0 = 0) in a matched and an unmatched arm. All arms share the paper's prior, drive, and rollout under one training recipe, and the reference is the default configuration with A0A_0, retrained under this recipe. The matched lag-only decoder keeps decoder parameters within 1% and has the same solved-gain ceilings (0.557 operational and 0.985 full budget, Secs. 3.2–3.3 and Appendix A.3.1), so the stability bound is the same. The attenuation factor of the right-forearm row of Sec. 4.2 is set per draw so that the forearm Z-rotation ROM over frames 6–180 is 0.80 ± 0.02 of the base. Responses over the same frames are averaged over the draws meeting this target.

Configuration (3 seeds each)FID per seed (Appendix A.2.2; all 95% intervals overlap)Recognition accuracyMatched draws (of 2,400)Cross-joint response, relative ROM (first row = 1)Cross-joint response, trajectory RMS (first row = 1)
With A0A_0 (default configuration)0.373 / 0.380 / 0.4640.87–0.881,66311
Lag-only, matched0.348 / 0.372 / 0.3470.86–0.871,2851/3.41/7.2
Lag-only, unmatched (A0A_0 removed only, fewer parameters, lower ceiling)0.467 / 0.373 / 0.3530.87–0.881,7091/8.01/15.9

At matched target effect, the cross-joint response is smaller by a factor of 3.4 (relative ROM) and 7.2 (trajectory RMS) in the matched lag-only decoder than in the default configuration, and lower in both lag-only arms in all nine seed pairings. The lagged operators of the matched lag-only decoder do not take over the cross-joint coupling of A0A_0 (92–94% of their weight stays within-channel, against 94% in the paper's models, Sec. 4.1).

As defined in Sec. 3.1, the response magnitude measures how much an action depends on the edited rule. These results show that, in the default configuration and with no resolvable difference in FID, the contemporaneous DAG carries forearm coordination that the lag-only decoder does not replace. The contemporaneous mechanisms thus add an explicit, editable rule for direct cross-joint coupling within the stability bound.

The gain constraints also matter for intervention stability. As reported in Sec. 4.4, the lagged cross-joint share of the paper's configuration rises from 5.5% to 70.6% without gain constraints, but a fivefold amplification request then produces numerical explosions in 47.2% of 360-frame rollouts, against none with the constraints (Table 4).

Relation to GOM. MotionSCM builds on GOM's central idea, explicit time-varying dependencies between body parts (Olivas-Padilla et al., 2024), and extends it with causality. We treat these dependencies as structural assignments of a conditional generator, define which interventions on them are admissible, and, as the review notes, limit their causal interpretation to the learned generator (Secs. 3.1, 3.3, and 5). GOM fits its coefficients to a recorded motion or decodes them from a recurrent state fed by the preceding postures, where editing a coefficient also changes the postures that set later ones. Our operators and drive are fixed by a latent trajectory sampled before rollout, independent of the generated state (Sec. 3.1), so one rule can be replaced while the latent, drive, seed frames, and all other rules stay fixed. The gain constraints bound the effect of every admissible edit (Sec. 3.2, Table 4), which GOM does not guarantee. Finally, GOM couples joints only across frames, whereas our contemporaneous DAG adds same-frame coupling, which the control above isolates.

AuthorsOpenReview ↗

Response to W2 and Q2 (3/4): Robustness of the coordination interpretations and sensitivity to the drive rank

Retraining and drive rank. We tested how consistently retrains of the paper's configuration recover its per-action rankings (W2) and how these rankings change with the drive rank (Q2). An independent three-seed retrain repeats this configuration and is the reference with A0A_0 in our reply to W1 and Q1. The R=16R = 16 and R=64R = 64 retrains (three seeds each) differ from it only in the rank of the jointly learned drive, which we set to R=32R = 32 in Appendix A.2.3. We compare per-action rankings of the Sec. 4.2 response with those of the paper's three R=32R = 32 models, using three-seed means for the independent retrain and single models for the drive-rank retrains. For single models, agreement is the reproducibility statistic of Appendix A.3.2 cited in W2, the mean Spearman correlation over seed pairs at the fixed strength of Sec. 4.2. Because training seeds alone already change single-model rankings, DD is the agreement across configurations minus the mean of both configurations' own-seed agreements, so a 95% interval of DD below 0 marks rankings that differ more across configurations than across seeds.

The independent retrain reproduces the three-seed-mean ranking of the Sec. 4.2 per-action response at Spearman 0.755 [0.482, 0.880] over the 25 actions, so this ranking is not specific to the paper's three trained models. Mean FID is 0.469, 0.406, and 0.373 for the R=16R = 16, 32, and 64 retrains, with all 95% intervals overlapping, and recognition accuracy is 0.82–0.86, 0.87–0.88, and 0.88.

Rankings, 25 actionsAgreement with the paper's R=32R = 32 modelsOwn-seed agreementDD [95% CI], verdict
R=16R = 160.5590.586−0.026 [−0.162, +0.080], no detected change
R=64R = 640.5410.495+0.001 [−0.102, +0.084], no detected change
Paper's R=32R = 32 models, same evaluation (new prior draws)0.585

The same comparison detects the ranking change between our ordering rule and the random permutation in our reply to Q3 (DD = −0.339 [−0.592, −0.107]), and both drive-rank intervals exclude a change of that size. These results show that, although the drive is learned jointly with the coupling, halving or doubling its capacity leaves single-model rankings without detected change.

Horizon. The horizon result cited in W2 is the reason we use k=5k = 5. As reported in Sec. 4.4 and Appendix A.3.2, longer lag windows lower FID without making action-specific responses more reproducible.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for Reviewers and Area Chair

Dear Area Chair and Reviewers,

We thank all the reviewers for the thoughtful reviews. In this work, we propose MotionSCM, a human motion generator that introduces a novel form of control by making cross-joint coordination mechanisms directly intervenable through structural causal assignments. We appreciate the reviewers' recognition of our main contributions: (1) the usefulness of explicitly intervenable cross-joint coordination mechanisms, (2) the theoretical rigor of stability guarantees for admissible interventions, and (3) the empirical effectiveness of coordination analysis alongside competitive generation quality.

We have conducted additional experiments and updated the manuscript (changes marked in blue). We believe these results successfully address the main concerns raised in the reviews and further strengthen our contributions.

ConcernAdditional evidenceKey result
Added value beyond GOM (kyrM)Matched lag-only control (Sec. 4.4, App. A.3.4)3.4× larger ROM and 7.2× larger RMS responses with the contemporaneous DAG; overlapping FID intervals
Reliability of coordination findings (kyrM, hDF9)Retraining, drive-rank and ordering controls, known-generator benchmark (App. A.3.4)Ranking reproduced (Spearman 0.755); no detected change beyond seed variation under tested drive ranks or alternative coefficient-based ordering
Editing effectiveness under matched conditions (hDF9)Matched-source latent editing comparison (Sec. 4.2, App. A.3.3)96% success; 28–89× lower off-target ratio on paired successful edits
Persistence of effects beyond frame 25 (NMW2)Full-sequence analysis (App. A.3.1)Effects persist to frame 180 as mean shifts in unedited channels

As we have not yet heard back from some reviewers, we would welcome any remaining questions before the deadline.

We hope these additional results and the revised manuscript will be taken into account in the final evaluation. Thanks again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.