acceptodds
Under review as a conference paper at ICLR 2027

Recency Forcing: Scaffolding Temporal Dependency for Long Autoregressive Video Generation

Abstract

Autoregressive (AR) video generation is causal by construction, yet how a generated frame depends on individual past context frames remains poorly understood, especially across the denoising steps of few-step distilled models. We introduce the positional response , a perturbation-based sensitivity measure quantifying how strongly predictions depend on a context frame at temporal distance and denoising step . Measured across distilled AR video diffusion models, reveals a consistent two-axis structure: influence decays rapidly with distance, while later denoising steps draw on a broader temporal window than earlier ones. Motivated by this, we propose Recency Forcing, applying a timestep-dependent Temporal Response Bias (TRB) to attention logits that scaffolds attention to match the measured response. This also mitigates KV eviction mismatch, where distant frames must be evicted from the KV cache during long-horizon inference despite being available in training: since TRB already drives their influence toward zero, eviction has little impact, letting short-clip-trained models generate substantially longer videos. We further introduce Biased Attention Reparameterization (BAR), an exact reformulation compatible with FlashAttention at negligible overhead. Recency Forcing supports both training-free and training-based settings, and experiments on VBench and VBench-Long show consistent long-horizon quality improvements with negligible inference overhead.

Primary area:
generative models
Keywords:
video generation, autoregressive model

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_r7KTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper measures how much an autoregressive video model's output depends on a past frame at each temporal distance and denoising step. Based on the pattern, it adds a step-dependent bias on attention to past frames (TRB), implemented through a FlashAttention-compatible rewrite (BAR). Used training-free or with short retraining, it improves long-video quality over prior methods at little extra cost.

Critical strengths
  1. A simple diagnostic that directly motivates the method: older frames matter less, and later denoising steps rely more on older frames.
  2. Practical: works with or without retraining, generates far beyond the training length, and BAR is much faster than adding the bias directly.
Critical weaknesses
  1. Key effects are not isolated. On short videos the gain is not separated from retraining, and the paper says the bias barely matters within the training window while also claiming it improves generation. There is no ablation with the same bias at every step, although step dependence is the core design.
  2. Limited generality and a weak user study. Only one small backbone is used. The user study reports only percentages, without significance or agreement.
Questions for authors
  1. How does an equally retrained model without TRB, and a step-independent bias of the same strength, perform?
  2. Can you report significance and per-pair counts for the user study? In Figure 7, color consistency and subject consistency have exactly the same preference rates against both Self Forcing and LongLive; is this correct?
  3. Does the pattern hold on other backbones?
7 comments
AuthorsOpenReview ↗

W2 / Q2. Weak user study and significant test - Same preference rates in color consistency and subject consistency

W2 / Q2. Weak user study and significant test - Same preference rates in color consistency and subject consistency

Justification on user study. We conduct a user study with 45 participants, collected 3384 responses. For context, prior work in this field typically uses 20–30 participants, so our user study is at or above the common scale.

We ran a two-sided binomial test on the collected responses. We note that statistical stablility in a 2AFC study is driven by the number of pairwise comparisons rather than the number of participants. Across all four aspects, Recency Forcing is preferred at a statistically significant rate:

Recency-Forcing vs Self-Forcing

AspectPreference95% CIpp (two-sided)
Overall quality94.7% (144/152)[90.8%, 98.0%]<0.001<0.001
Color consistency94.1% (143/152)[90.1%, 97.4%]<0.001<0.001
Dynamic motion92.1% (140/152)[87.5%, 96.1%]<0.001<0.001
Subject consistency94.1% (143/152)[90.1%, 97.4%]<0.001<0.001
All aspects93.8% (570/608)[91.8%, 95.6%]<0.001<0.001

Recency-Forcing vs Rolling-Forcing

AspectPreference95% CIpp (two-sided)
Overall quality75.9% (107/141)[68.8%, 83.0%]<0.001<0.001
Color consistency74.5% (105/141)[67.4%, 81.6%]<0.001<0.001
Dynamic motion73.0% (103/141)[66.0%, 80.1%]<0.001<0.001
Subject consistency72.3% (102/141)[64.5%, 79.4%]<0.001<0.001
All aspects73.9% (417/564)[70.4%, 77.5%]<0.001<0.001

Recency-Forcing vs LongLive

AspectPreference95% CIpp (two-sided)
Overall quality66.9% (87/130)[58.5%, 74.6%]<0.001<0.001
Color consistency64.6% (84/130)[56.2%, 73.1%]0.0010.001
Dynamic motion73.8% (96/130)[66.2%, 80.8%]<0.001<0.001
Subject consistency64.6% (84/130)[56.2%, 73.1%]0.0010.001
All aspects67.5% (351/520)[63.5%, 71.5%]<0.001<0.001

Same preference rates in color consistency and subject consistency.

This is true, the preference rates for color consistency and subject consistency are identical in the comparisons against Self Forcing and LongLive. The complete preference rates are reported above. One possible explanation is that these two criteria may be correlated, as color drift can coincide with changes in subject appearance.

AuthorsOpenReview ↗

W2 / Q3. One small backbone was used and Temporal Response Analysis on other backbones

W2 / Q3. One small backbone was used and Temporal Response Analysis on other backbones

Justification on one backbone. We acknowledge the limited backbone coverage in Appendix G. Limitations. However, Wan2.1-1.3B is the common testbed used by all baselines in this line of work (Self-Forcing, Causal Forcing, Rolling Forcing, LongLive, Reward Forcing), primarily for compute feasibility and fair comparison.

We further conduct our temporal response analysis on the Wan2.2-5B few-step autoregressive model released by LongLive2.0 [1], extending the analysis beyond Wan2.1-1.3B. To the best of our knowledge, this is the only available few-step autoregressive distilled student based on a backbone other than Wan2.1-1.3B. The released model uses a 32-frame attention context window, generates 8 frames per chunk, and runs 4 denoising steps, meaning only 4 chunks in attention window. Thus, we intentionally extend the model's attention window to 64 frames (8 chunks) for measuring R~\widetilde{R}.

The normalized positional response R~\widetilde{R} plot: LongLive2.0 5B's R plot

The results exhibit the same pattern reported in the main paper: monotonic ordering, steep decay, and denoising-step dependence, providing additional evidence that these trends extend beyond Wan2.1-1.3B. Interestingly, we observe a slight upward rebound at the tail of the curve (Δ=7\Delta=7) for the LongLive2.0 5B model. We hypothesize that this behavior may be attributed to the use of Attention Sinks along with the model's training on long-horizon videos.

We have added this new analysis in Appendix B. Positional Response on Wan2.2-5B in our revised paper.


References

[1] Chen, Yukang, et al. "LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation." arXiv preprint arXiv:2605.18739 (2026).

AuthorsOpenReview ↗

W1 / Q1. An equally retrained model without TRB and step-independent bias

W1 / Q1.1. "The bias barely matters within the training window"

Clarify: Adding TRB improves short video generation. As shown in Table 3 of the main paper, our method achieves the highest overall score on VBench for short videos. Thus, the bias does not "barely matter" within the training window.

We acknowledge that our original phrasing contradicted the results, and we have revised the paper in Section 4.1:

Old: "Our method achieves the best overall score on VBench, demonstrating that integrating TRB introduces no degradation in short-video quality - the bias has negligible effect when all context frames fit within the training window."

New: "Our method achieves the best overall score on VBench, demonstrating that integrating TRB actively benefits short-video generation."

W1 / Q1.2. An equally retrained model without TRB and step-independent bias

The model without TRB is Causal Forcing baseline, reported in Tables 3 and 4 in the Main paper for short- and long-video generation. Both methods use the same training dataset and initialization from the Causal Forcing stage-2 causal ODE checkpoint, as detailed in Appendix D Implementation Details.

The only differences between our training settings and original baseline's settings are training iterations and batch size. To additionally control the settings, we retrain the Causal Forcing baseline (without TRB) for 3,500 iterations and batch size 4 - matches the training setting of Recency Forcing (same dataset, same initialization,...). The evaluation protocol is the same on both VBench and VBench long reported in the main paper.

Short-video generation: VBench, 5 seconds

ModelTotal ↑Quality ↑Semantic ↑
Causal Forcing (original)84.0484.5981.84
Causal Forcing (retrained, 3,500 iterations, batch size 4)81.8482.8277.90
Recency Forcing85.0886.2080.59

Long-video generation: VBench-Long, 60 seconds

ModelQuality ↑Background ↑Subject ↑Motion ↑Dynamic ↑Aesthetic ↑Imaging ↑Temporal ↑
Causal Forcing (original)78.9895.5195.3996.9465.1951.3563.4495.21
Causal Forcing (retrained, 3,500 iterations, batch size 4)79.2894.8094.3496.0668.3457.4066.9593.75
Recency Forcing84.02 (+4.74)96.4997.6898.0075.4861.3170.5096.32

Result: Under the same training setting, adding TRB improves every short-video score, and Total rises from 81.84 to 85.08 (+3.24). On 60-second videos, TRB consistently improves all dimensions, and Quality rises from 79.28 to 84.02 (+4.74). Since every other factor is held fixed, these improvements come from TRB and not from differences in training.

To isolate denoising-step dependence, we provide a comparison between step-dependent (Recency Forcing) and step-independent variants, using our best-performing exponential bias by fixing tdenoise=1t_{\mathrm{denoise}}=1 throughout training and inference, while keeping the other settings unchanged.

Short-video generation: VBench, 5 seconds

Bias configurationTotal ↑Quality ↑Semantic ↑
Step-independent83.7684.6880.06
Step-dependent (original Recency Forcing)85.0886.2080.59

Long-video generation: VBench-Long, 60 seconds

Bias configurationQuality ↑Background ↑Subject ↑Motion ↑Dynamic ↑Aesthetic ↑Imaging ↑Temporal ↑
Step-independent82.7596.4197.8298.0861.9659.8769.9096.38
Step-dependent (original Recency Forcing)84.0296.4997.6898.0075.4861.3170.5096.32

Result: Our original step-dependent TRB design form outperforms the step-independent variant on both short and long horizon VBench. On short VBench, Total Score improves by +1.32, Quality by +1.52, and Semantic by +0.53. On 60-second videos, it improves Quality by +1.27 (84.02 vs. 82.75), and the gap comes from Dynamic (75.48 vs. 61.96, +13.52).

Reviewer_c45YOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Recency Forcing for extending short-clip-trained autoregressive video diffusion models to longer generation horizons. The authors introduce a perturbation-based positional response that measures prediction sensitivity to context frames across temporal distances and denoising steps. Motivated by the observed patterns, Temporal Response Bias attenuates historical context while preserving recent and global sink frames. Biased Attention Reparameterization implements this bias through FlashAttention under a shared temporal-index condition. The paper reports improvements on 60-second video generation in both training-free and training-based settings, supported by VBench evaluations, a human preference study, and an extension to frame-wise generation.

Critical strengths
  • The paper addresses a practically relevant problem with a simple intervention.

  • The evaluation includes useful comparisons with context truncation, learned bias, and several parametric bias families. The human preference study and frame-wise results provide additional evidence of practical effectiveness beyond the main aggregate scores.

  • BAR is a useful implementation contribution. Its algebraic equivalence is clearly explained under the stated query-grouping condition, and the reported end-to-end latency increases only from 4.912 to 5.076 seconds for five-second generation.

Critical weaknesses
  • The benefit of the response-guided, timestep-dependent design is not sufficiently isolated. The biggest concern for me is the connection between the measured positional response and the proposed attention modification. The response already reflects the model’s existing attention, value representations, and subsequent nonlinear transformations. Applying a similarly shaped attenuation to attention logits does not establish that the resulting dependency will match that response or improve generation for the proposed reason.

  • The positional-response experiment does not directly establish the KV-eviction explanation. In Equation (4), replacing one frame with a frame from another rollout of the same prompt can still create an incompatible temporal context, even if the replacement itself is a plausible frame. Therefore, this procedure does not guarantee the absence of out-of-distribution perturbations claimed in Section 3.2. Moreover, sensitivity to replacement with the remaining history fixed differs from sensitivity to removing cached keys and values, and it does not capture an earlier frame’s effects mediated through subsequently generated frames.

  • The reported gains do not fully resolve the trade-offs between visual quality, motion, and conditional alignment. Dynamic degree assesses the extent of motion, so an increase alone does not establish more natural motion. The paper itself demonstrates this limitation through the truncation baseline in Table 5. This distinction is consistent with the VBench metric definition. Conversely, in the frame-wise setting, dynamic degree decreases from 79.71 to 60.71 while quality improves; this could reflect removal of unstable motion or reduced useful motion, and the table does not distinguish them.

  • The decay schedule contains inconsistencies that affect interpretation and reproducibility. Equation (8) implies that increasing γbase\gamma_{base} makes the bias less negative and therefore weakens attenuation. Section 4.2 states the opposite: small values insufficiently suppress history, while large values over-suppress it. Please reconcile this explanation with the implemented schedule and Table 6.

Questions for authors

See the weaknesses.

5 comments

W3 / Q3. Dynamic degree decreases in Frame-wise setting

We agree that Dynamic degree alone cannot tell whether unstable or useful motion was removed (Appendix E, Generalization to the frame-wise setting). Instead, unstable motion or useful motion is separated by other motion-related metrics and the usual range of Dynamic degree:

(1) Temporal Flickering and Motion Smoothness improve. Our frame-wise setting has high Temporal Flickering score 96.56 and Motion Smoothness score 98.18 (higher is better, Table 8 Appendix E, generalization to the frame-wise setting). Temporal Flickering measures frame-to-frame consistency, and Motion Smoothness measures whether motion is smooth and plausible. Abrupt motion, such as jitter or sudden jumps between frames, lowers both scores.

(2) Dynamic degree is still above the usual range. Dynamic degree 60.71 in Recency Forcing frame-wise setting is well above Self Forcing (34.50), Rolling Forcing (34.68), and LongLive (42.85) in Table 4 of the main paper (60-second evaluation).

Together, this suggests lower Dynamic degree might come from reducing unstable motion, not from a loss of motion overall. We note that these metrics only indirectly separate unstable from useful motion, we do not claim that no useful motion is lost.

Importantly, frame-wise is not our primary setting, it reuses the chunk-wise hyperparameters without any re-tuning. Further tuning could improve the balance between Dynamic degree and quality.

W4 / Q4. γbase\gamma_{base} opposite explanation

We thank the reviewer for pointing this out. The explanation in Sec. 4.2 ablation study, paragraph “Effect of decay strength γbase\gamma_{\mathrm{base}}” (L467-469) was opposite. We have corrected it and added details for clarity in the revised paper:

Old: "...small values insufficiently suppress distant frames, while large values over-suppress history..."

New: "...small γbase\gamma_{\text{base}} values over-suppress history, while large γbase\gamma_{\text{base}} values insufficiently suppress distant context..."

W1 / Q1. Connection between the Positional Response Analysis and the attention modification

1. Clarification on what RR measures and our motivation. RR measures how strongly the prediction responds to the content of a context frame - not a measurement for attention weight on context. It reveals the natural recency, timestep-dependent bias of Few-step AR models. And our motivation is how to use this information as a prior knowledge to improve Few-step AR video generation. We do not try to make the re-trained model matches the same response (Sec. 1 Introduction L73-L79).

The connection between RR and attention modification. The question is how to exploit this prior knowledge (the relationship between the current denoising frame and context frames), to improve Few-step AR video generation → Self Attention is a straightforward design choice. In DiT architecture, Self Attention is the only place where the current denoising frames interact with history frames. Other DiT layers like MLPs process each token independently, and cross-attention interacts with text tokens.

2. Why this helps improving video generation. Content response and attention mass might disagree: a distant frame can have near-zero response on content yet still receive attention mass.

  • On short horizon: Because the content concentrates on a narrow range of context, our attention attenuation design (TRB) scaffolds the attention mass toward this informative range, improving short horizon video generation.
  • On long horizon: During long rollouts, distant frames are evicted from the KV cache, so the attention distribution at inference differs from training. TRB helps mitigating this: it drives distant-frame attention mass toward zero before eviction, evicting that frame might has little effect on the prediction.

3. And we show this improvement through empirical experiments, not only by intuition:

  • Short videos evaluation (Table 3 Main paper): Best overall score on VBench.
  • Long videos evaluation (Table 4 Main paper): Highest quality score on VBench-Long with training-based setting.
  • Ablation the effect of applying TRB (Table 5 Main paper): Performance consistently improves when applying TRB (regardless of the decay function used).

W2 / Q2. RR does not establish KV-eviction explanation / sensitivity to KV eviction / replacement does not eliminate OOD

Clarification on the purpose of RR and frame replacement. Our proposed positional response RR does not aim to establish the KV-eviction explanation, nor is it designed to measure sensitivity to KV eviction. Rather, RR only serves a diagnostic that measures how strongly the prediction responds to the content of a context frame at temporal distance Δ\Delta.

Regarding the perturbation method, replacement is not designed to eliminate out-of-distribution perturbations, rather, replacement is the reasonable perturbation for isolating the effect of the context frame's content.

And we acknowledge that our word choice "avoid" (Sec. 3.2 Analysis, L203-L204) was inaccurate, and we have revised to clarify that replacement is intended to "reduce" out-of-distribution perturbations compared with additive noise - an extreme pertubation, rather than eliminate OOD.

Old: “We use replacement rather than additive noise to avoid introducing out-of-distribution perturbations.”

New: “We use replacement rather than additive noise to introduce less extreme perturbations and reduce the risk of out-of-distribution.”

Please read the authors' rebuttal and post your comments on your raised concerns.

AC

Reviewer_DsNDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce a perturbation-based analysis that measures how replacing individual past frames affects model predictions across denoising steps, finding that distant frames have less influence and that later denoising steps rely on a broader temporal context. Based on this observation, they propose Recency Forcing, which applies a timestep-dependent attention bias to suppress distant frames before KV-cache eviction, improving generation quality up to 60 seconds for models trained on 5-second videos.

Critical strengths
  • The paper asks a significant question: How does the generation of a single frame depends on individual context frames? This simple but overlooked question shapes the entire motivation of the paper.

  • The paper provides an interesting perspective on how autoregressive video diffusion models utilize temporal context by the perturbation-based positional response metric to measure how individual historical frames influence model predictions at different denoising steps.

  • The proposed Temporal Response Bias (TRB) is straightforward to implement and can be applied in both training-free and training-based settings.

Critical weaknesses

First of all, the qualitative results in supplementary material are very limited. 1 or 2 prompts in each experiments, surely not enough for assessing the quality of long video generation framework.

Second, the positional-response analysis does not disentangle the effects of temporal RoPE from the model's actual temporal dependency:

In Figure 2, the authors observe a sharp decay in positional response as temporal distance increases and interpret this as evidence that distant frames have less influence on the model's predictions. However, since the underlying model uses temporal RoPE, I wonder whether this decay is largely influenced by the positional encoding itself rather than the actual importance of historical information.

I believe this distinction is important, as the proposed Temporal Response Bias (TRB) is directly motivated by this observation. An ablation analyzing how the positional-response curves change under different temporal RoPE configurations would help clarify whether the observed decay reflects temporal dependency or is primarily induced by positional encoding.

Questions for authors

In Figure 2, the authors observe a sharp decay in positional response as temporal distance increases. Could this decay be largely influenced by temporal RoPE rather than reflecting the actual importance of historical frames? Have the authors considered analyzing the positional response under different RoPE configurations to disentangle the effect of positional encoding from the model's temporal dependency?

2 comments

W1 / Q1. Limited qualitative results

The supplementary material is limited to 100 MB, so we prioritized long videos with minimal compression and could only include a few prompts.

To address this, we have added many more results to our anonymous website: https://results-36644x.anonymous26qwe.workers.dev. The new results include:

  • Longer videos: videos of around 4 minutes, which test stability over a much longer horizon than the 60-second videos in the main paper.
  • Both settings: more prompts for both the training-based and training-free variants.

W2 / Q2. Influence of temporal RoPE

We thank the reviewer for this insightful question. Following the reviewer's suggestion, we ran new experiments showing that the steep decay is not an artifact of temporal RoPE: all three properties reported in the main paper persist under modified temporal RoPE configurations.

Setup. We assign all context frames with Δ ≥ 2 the same temporal RoPE index c (c = 0 or c = i−2). They then all have the same RoPE offset from the current frame, so any difference in R among them cannot come from positional distance. The current frame and the most recent frame (Δ = 1) keep their original indices, since changing them makes the rollout fail.

Plots. The shape of R is easiest to judge visually, so we show the normalized positional response R (normalized by R at Δ = 1, for each denoising step) under the original configuration and both modified ones. The plots are also in Fig. 7 (Appendix A, our revised paper).

Results. In all three plots, R shows the three patterns discussed in the main paper (Sec. 3.2, L212–L215): steep decay, the same ordering across denoising steps, and divergence at large Δ.

  1. Decay persists under different temporal RoPE configurations. Among frames that all share the same RoPE offset (Δ ≥ 2), R falls by roughly 6–12× from Δ = 2 to Δ = 5, and decreases strictly with Δ from 2 to 9 in every setting.
  2. The steepest drop at Δ = 2 does not depend on RoPE. R(Δ = 2) ≈ 0.08 in all settings.
  3. The ordering across denoising steps is preserved. The curves remain ordered by denoising timestep at all distances.

Important note. Our original analysis studies the few-step AR model under its original video-generation configuration, and does not separate the contributions of RoPE and context dependencies. The new experiments assess the influence of temporal RoPE further, but they do not fully disentangle it, because the model no longer runs in its original configuration. We thank the reviewer for motivating this analysis, which strengthens our original one. More implementation details and results are in Appendix A of the revised paper.

Please read the authors' rebuttal and post your comments on your raised concerns.

AC

Comments on the paper

2 comments

Thank you for your reply. Our main point is that both RoPE and frame content influence the shape of RR, rather than RR being determined solely or predominantly by RoPE. Nowhere in the paper do we claim that RR depends exclusively on either frame content or RoPE.

Specifically, Fig. 2 in the main paper investigates how the shape of RR changes when varying frame content while keeping RoPE fixed. Conversely, the figure provided here ((https://anonymous.4open.science/api/repo/recency_forcing_R_plots/file/R-pos-plot.jpg) examines how RR changes when varying RoPE while keeping frame content fixed.

Together, these two figures demonstrate that RR is influenced by both RoPE and frame content, challenging the reviewer's earlier suggestion that RoPE is the dominant factor. In particular, for a near-frame positional change (Δ=1→2\Delta=1\rightarrow2), RposR_{\mathrm{pos}} decreases only from approximately 1.01.0 to 0.60.6, whereas varying frame content causes a much sharper decrease in RR, from approximately 1.01.0 to 0.080.08.

Importantly, we do not directly use the numerical values from these plots to construct the prior or scaffolding function in TRB. Instead, these observations motivate our design of the TRB function, whose effectiveness is empirically validated. Notably, TRB improves performance even for short-video generation, as shown in Tab. 3 of the main paper, where the score increases from 84.04 (Causal Forcing) to 85.08 (Recency Forcing).

To the Area Chair and All Reviewers,

We sincerely thank the Area Chair and reviewers for their constructive feedback and careful evaluation. We are encouraged by their recognition of the importance of our research question, the value of our diagnostic perspective, and the effectiveness of our approach to long-video generation.

Revision tracking: We have highlighted all revisions in red in the updated submission PDF to help the Area Chair and reviewers easily identify the changes.

To assist the Area Chair, we summarize the main concerns, our responses, and the revisions made to strengthen the paper. We have focused on clarifying the interpretation of our diagnostic, isolating the contribution of TRB, and providing additional evidence for robustness and generality.

Reviewer DsND

Initial Concerns:

  1. Limited qualitative results: Only 1–2 prompts were included in the supplementary material.
  2. Temporal RoPE entanglement: The decay in Positional Response (RR) might be caused by temporal RoPE rather than temporal dependencies.

Our Response:

  1. Expanded qualitative results: We shared an anonymous website with more and longer videos in both training-based and training-free settings.
  2. RoPE ablations: Fixing context frames' temporal RoPE indices shows that the decay is not merely a RoPE artifact (Appendix A). A complementary experiment varying RoPE while fixing frame content shows that both factors influence the shape of RR, motivating our Temporal Response Bias (TRB). Importantly, the observed response curves guide the design rather than serving as numerical targets for the bias.

Reviewer c45Y (Score: 2, Weak Reject → 3, Weak Accept)

Initial Concerns:

  1. RR and attention modification: A similarly shaped attention bias need not reproduce RR or explain performance gains.
  2. KV-eviction and OOD claims: RR does not prove KV-eviction, and frame replacement cannot eliminate all OOD effects.
  3. Frame-wise motion: Does lower Dynamic Degree remove unstable or useful motion?
  4. Inconsistency: The explanation of γbase\gamma_{\text{base}} conflicted with Eq. (8).

Our Response:

  1. Clarified RR: It measures sensitivity to context-frame content, not attention weights, and serves as a design prior rather than a curve to match. Tables 3–5 validate the gains empirically, including improvements within the short-video training window.
  2. Refined claims: We clarified that RR is diagnostic, not proof of KV-eviction (L142–144), and that frame replacement reduces, rather than eliminates, extreme OOD perturbations (Sec. 3.2).
  3. Clarified motion results: Dynamic Degree remains above the baselines, alongside strong temporal consistency and smoothness, suggesting reduced unstable motion. Frame-wise generation is included to test generalizability, not as our primary setting.
  4. Fixed the typo: We aligned the γbase\gamma_{\text{base}} explanation in Sec. 4.2 with Eq. (8).

Final Feedback: The reviewer recognized RR as a design prior and, despite a remaining theoretical question about attention redistribution under KV-eviction, stated that "the paper is worth reading" and "a new perspective to extend generation context should be valued." They raised their score from 2 to 3 (Weak Accept).

Reviewer r7KT (Initial score: 3 - Weak Accept)

  1. Concern 1 (Effects Not Isolated & Short-Video Claim): The reviewer questioned whether gains came from retraining or step dependence and asked why the retrained baseline underperformed the original Causal Forcing on short videos.

    • Our Response: An equally retrained baseline without TRB isolates TRB's contribution (+3.24 short-video VBench Total; +4.74 long-video Quality). A step-independent ablation confirms additional gains from step dependence (+1.27 long-video Quality; +13.52 Dynamic Degree). The short-video baseline gap is attributable to a 16× smaller batch size (4 vs. 64) and noisier DMD gradients; Recency Forcing still outperforms both baselines under this budget.
    • Status: [RESOLVED]
  2. Concern 2 (User Study Rigor): The reviewer requested exact counts, significance tests, and an explanation for identical color/subject consistency rates.

    • Our Response: We reported per-pair counts (3,384 responses), 95% confidence intervals, and two-sided binomial tests (pp-values), and explained the identical rates through correlated color/subject drift.
    • Status: [RESOLVED]
  3. Concern 3 (Backbone Generality): The original evaluation used only Wan2.1-1.3B.

    • Our Response: We extended evaluation to Wan2.2-5B (LongLive2.0). Applying training-free Recency Forcing without re-tuning yielded improvements (+0.59 Quality). The reviewer acknowledged that these 5B results were what they asked for, but due to rebuttal time constraint and last minute effort, only training-free on version of Wan2.2-5B is evaluated .
    • Status: [PARTLY ADDRESSED]

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.