acceptodds
Under review as a conference paper at ICLR 2027

Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation

21 views

Abstract

Autoregressive video diffusion models enable video generation by sequentially denoising video chunks conditioned on preceding context. Beyond producing successive chunks, a key capability expected from this paradigm is to model how visual scenes evolve over time, which is also central to world modeling. However, teacher forcing, together with the strong temporal redundancy of video, makes current-chunk denoising susceptible to shortcut learning, allowing the model to exploit local visual continuity without sufficiently capturing the dynamics of video evolution. To bridge this gap, we introduce Future Chunk Prediction (FCP), which augments autoregressive training with future-chunk supervision that is less directly supported by local visual continuity, encouraging the model to better capture how visual scenes evolve over time. We evaluate physical plausibility as a challenging manifestation of video evolution modeling. Our method achieves superior physical plausibility over standard teacher-forced autoregressive training with a lower overall training compute budget, highlighting the effectiveness and efficiency of future-chunk supervision for modeling visual evolution. Intriguingly, FCP also induces more spatiotemporally structured representations, suggesting a promising direction toward predictive representations for world modeling.

Primary area:
generative models
Keywords:
Generation, Autoregressive, Diffusion

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_2uNEOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FCP adds future-chunk prediction losses to autoregressive video diffusion, encouraging current representations to support later predictions. Parallel auxiliary heads condition on clean/noisy features and are discarded at inference. Wan 1.3B/5B experiments examine physical plausibility and auxiliary design choices.

Critical strengths
  1. Consistent gains without added inference cost. FCP improves Physics-IQ Verified from 27.6 to 30.9 on Wan2.1-1.3B and from 30.4 to 33.8 for I2V and 40.6 to 44.4 for V2V on Wan2.2-5B (Tables 2 and 4). Table 3 reports gains on complementary benchmarks.

  2. Relevant design ablations. Tables 2, 4, 6, and 7 examine prediction structure, source features, text, offsets, and feature injection. On 5B, parallel prediction exceeds the chained variant by 2.8 I2V and 4.5 V2V IQ points.

Critical weaknesses
  1. The shortcut explanation rests on indirect evidence. Sections 1 and 5.5 link the benefits of FCP and parallel prediction to reduced reliance on local continuity and intermediate future targets. The ablations show performance gains but do not isolate these mechanisms from other auxiliary-training effects. Figure 3 shows two successful examples; the distribution of improvements, regressions, and remaining physical errors across scenarios is unclear.

  2. Uncertainty and compute accounting are incomplete. Appendix A.3 reports four-generation-seed means without dispersion or paired intervals, including for the 0.8-point 1.3B parallel/chained gap. Figure 4 derives 42.6%/50.7% compute savings over full-context FCP from exponential fits to non-monotonic scores. The FLOPs estimator, fitting procedure, and threshold-crossing sensitivity are unspecified (§5.5).

Questions for authors
  1. Which physical scenarios improve or regress, and which errors remain? Report paired per-scenario changes, representative failures, per-seed scores, and confidence intervals for TF/FCP and parallel/chained comparisons, grouping camera views of each scenario.
  2. What evidence distinguishes reduced reliance on local continuity or intermediate future targets from other effects of the auxiliary objective and parallel architecture? Connect this explanation to the offset and component ablations in Tables 2 and 4.
  3. Does increasing the prediction horizon saturate or reverse the gains? Retaining information for distant predictions may compete for capacity with next-chunk fidelity. Extend the fixed-head offset comparison and vary auxiliary-loss strength at matched backbone capacity and training compute. Measure physical consistency and near-term fidelity to identify the best trade-off in the tested range.
  4. How are the main/auxiliary forward-backward FLOPs and Figure 4 fits computed? Report checkpoint values and the sensitivity of the claimed savings to fitting choices and observed checkpoint crossings.
6 comments
AuthorsOpenReview ↗

Response to Q1: Scenario-Level Results, Remaining Errors, and Statistical Reliability

The official Physics-IQ Verified score aggregation is equivalent to first averaging the three camera-view scores within each scenario and then averaging across scenarios. We report category-level mean scores using the official Physics-IQ categories below:

CategoryNumber of scenariosTFFCPDifference
Solid Mechanics3828.79132.126+3.335
Fluid Dynamics1529.93332.610+2.677
Optics817.64119.435+1.794
Thermodynamics335.97946.939+10.960
Magnetism214.38116.010+1.629
Overall6627.58930.883+3.294

FCP achieves higher mean scores than TF across all five physical categories, indicating that the gains are not limited to a particular type of physical phenomenon.

We will share additional video examples following our open-source release, alongside all checkpoints and training/inference code, to support community comparisons across scenarios and applications. Despite the overall gains over TF, our qualitative impression is that errors in rigid-body collisions, magnetism, and buoyancy remain readily noticeable to the human eye, potentially reflecting humans’ familiarity with rigid-body dynamics and limited training coverage of the latter two. We expect improved data coverage and curation to help address these limitations and will pursue these improvements in future versions.

To address the request for multi-seed evidence, we additionally provide representative aggregate IQ results for the TF/FCP and parallel/chained comparisons in Table 4, reported as mean ± standard deviation across four generation seeds:

ConfigurationPhysics-IQ Verified ↑ (Mean ± Std)
Teacher forcing27.6 ± 0.798
FCP — parallel30.9 ± 0.386
FCP — chained30.1 ± 0.893

The complete IQ mean ± standard deviation table covering all six configurations in Table 4 is provided in our response to Reviewer PWCw (Q6).

During training, we monitored the overall convergence curves of parallel and chained prediction. Parallel outperformed chained at 10k, 15k, and 20k steps, which represent only a subset of the checkpoints examined. Together with the scale-up results in Table 2, this supports a consistent advantage rather than an isolated checkpoint fluctuation. The standard deviations reflect the stochastic nature of autoregressive rollouts and the sensitivity of the evaluation metrics.

AuthorsOpenReview ↗

Response to Q2: Shortcut Mechanism and Parallel versus Chained Prediction

We conducted experiments with different prediction-offset configurations on the 5B model. The I2V Physics-IQ results are summarized below:

MethodPrediction offsetsPhysics-IQ Verified ↑
Teacher forcing—30.4
FCP — adjacent only{1}30.1
FCP — distant only{2,3,4,5}33.5
FCP — full configuration{1,2,3,4,5}33.8

Auxiliary supervision restricted to adjacent-chunk prediction does not improve physical adherence in this setting, whereas supervision of more distant chunks yields substantial gains. These results support our motivation that predicting more distant future evolution benefits model training, with FCP’s auxiliary objective serving as a practical implementation of this principle. For the rationale behind retaining the Δ=1\Delta=1 branch, please refer to our response to Reviewer PWCw (Q1).

We additionally evaluated the three-offset chained configuration, extending the parallel/chained comparison in Table 2. The complete 5B I2V comparison is shown below:

Number of offsetsPrediction structurePhysics-IQ Verified ↑
3Parallel32.1
3Chained30.7
5Parallel33.8
5Chained31.0

Chained prediction remains inferior even when the chain is shortened to three offsets. This rules out an explanation based solely on the longer five-offset configuration used in the original 5B comparison (§5.5, L427–429). We therefore focus our interpretation on potential leakage of adjacent-future information through the chain (Fig. 2a).

We have tried our best to align the predictor configurations, prediction targets, auxiliary-loss weights, and training settings within each offset set to isolate the connection pattern as closely as possible. In an end-to-end chained design, input dependencies and gradient paths are coupled through the sequential connections [1]. These factors cannot be varied independently within the present comparison without altering the chained design itself.

References

[1] DeepSeek-AI. DeepSeek-V3 Technical Report. arXiv:2412.19437, 2024. §2.2, Eqs. 21–25.

AuthorsOpenReview ↗

Response to Q3: Effect of Auxiliary Prediction Strength on Performance

We conducted an ablation study on the FCP auxiliary-loss weight. The Physics-IQ results are summarized below:

Method / FCP loss weightPhysics-IQ Verified ↑
Teacher forcing27.6
FCP, weight = 0.129.6
FCP, weight = 0.5 (default)30.9
FCP, weight = 2.028.1

Limited computational resources and job-queue delays during the rebuttal period precluded a finer-grained sweep. Nevertheless, the degradation at the largest tested weight shows that further strengthening FCP can become counterproductive. This is consistent with the need to balance auxiliary supervision against the primary generation objective: an excessively large auxiliary-loss weight can dominate optimization at the expense of the main task.

We further compared offset configurations on the 5B model under matched computation, using three auxiliary predictors in each case. The I2V results are shown below:

Prediction offsetsPhysics-IQ Verified ↑
{1,2,3}32.1
{1,3,5}33.1
{1,4,7}33.2

These results show that extending the prediction horizon can yield additional gains at the same computational budget. Although we observed no decline within the tested range, we expect gains to eventually saturate or reverse, as predicting increasingly distant futures without access to intervening context becomes increasingly underdetermined. Our exploration was limited by computational resources during the rebuttal and the scarcity of valid supervision pairs at larger offsets in 5 s training clips. For example, at Δ=9\Delta=9, only the first chunk of each sample can serve as a prediction source, making stable training more difficult.

We have not sufficiently explored combinations of loss weights and offsets to identify a joint optimum. Our resources are limited, and our experience suggests that suitable settings depend on the model, task, and training data. As an initial exploration, our work provides broadly applicable settings, while the optimal configurations for specific applications remain to be explored.

Reviewer_PWCwOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Future Chunk Prediction (FCP), an auxiliary objective for chunk-wise autoregressive video diffusion. Alongside teacher-forced denoising, lightweight predictors generate a chunk Δ steps ahead from the generator's intermediate hidden states of the current chunk, without access to the intervening chunks, and the loss is backpropagated into the generator. The predictors are discarded at inference and no pretrained encoder is needed. The authors convert Wan2.1-1.3B and Wan2.2-TI2V-5B into chunk-causal autoregressive models and evaluate mainly on Physics-IQ Verified, with WorldModelBench, PAI-Bench-G and PhyGround as secondary benchmarks. Under matched data and training steps, FCP improves Physics-IQ over teacher forcing at both scales and in both image-to-video and video-to-video settings.

Critical strengths
  1. Simple, training-only objective. FCP needs no external pretrained encoder and leaves autoregressive inference unchanged. It reaches higher Physics-IQ than teacher forcing at comparable estimated training compute (Fig. 4), and needs 42.6%/50.7% less compute than a full-history variant to reach IQ 29/30.
  2. Consistent controlled gains, reported transparently. With matched data and training steps, FCP improves Physics-IQ Verified from 27.6 to 30.9 at 1.3B and from 30.4 to 33.8 (image-to-video) and 40.6 to 44.4 (video-to-video) at 5B. Results are means over four seeds on all 198 samples, and the full training recipe (optimizer, learning rate, EMA, gradient clipping) is given in Appendix A.2. Gains also carry over to WorldModelBench (total 7.22 → 7.64) and PAI-Bench-G (domain score 81.13 → 82.38). PhyGround changes little (+0.03/+0.05), possibly because the VLM judge is insensitive to these differences, as the authors suggest.
Critical weaknesses
  1. The comparison to public models in Table 1 overstates FCP's contribution. Section 5.4 states that the 5B model "outperforms larger bidirectional models" (Wan2.2-14B 32.2, HunyuanVideo-1.5 33.4, Cosmos3-Nano 29.1). These models were not fine-tuned on the paper's physics-oriented training data (internal data plus WISA-80K). The paper's own teacher-forcing baseline already reaches 30.4, against 24.9 for base Wan2.2-5B, so most of the gap to public models comes from fine-tuning rather than from FCP (+3.4). The claim should be weakened, and the teacher-forcing row should be added to Table 1.
  2. Evaluation is limited to short continuations. All results are 5-second continuations within the training clip length (85 frames at 16 fps for 1.3B; 121 frames at 24 fps for 5B). For an autoregressive method motivated by world modeling, it remains untested whether FCP reduces or worsens error accumulation over longer rollouts, a limitation the authors acknowledge in Appendix C.3.
Questions for authors
  1. (Clarification, plus experiment) Eq. 5 defines FCP with Δ > 1, so that intervening chunks are excluded, while Eq. 6 uses Δ ≥ 1, and every configuration includes Δ = 1 ({1,2,3} for 1.3B, {1,…,5} for 5B). With Δ = 1, the predictor acts as an extra next-chunk head without text. Could you report ablations with Δ = 1 only and without Δ = 1 (e.g., {2,3} or {2,…,5})? These would show whether the gains come from far-future prediction specifically.
  2. (Experiment) Could you evaluate rollouts well beyond the training clip length (e.g., 2–4×), comparing FCP and teacher forcing on how quality or physical plausibility changes over time?
  3. (Experiment) The representation analysis (Figs. 5–9) uses hand-picked query tokens from a single layer (Appendix A.5). Could you provide quantitative evidence, such as correspondence or tracking accuracy, or linear probing over many tokens and layers, comparing FCP with teacher forcing?
  4. (Experiment) Section 4.1 states that sharing generator parameters between denoising and future prediction "degraded temporal coherence" in preliminary experiments. Could you report this comparison?
  5. (Experiment) The paper argues against representation-alignment methods because they rely on external pretrained encoders (§2), but never compares against them. Could you report a matched VideoREPA baseline (same backbone, data and training steps), since VideoREPA also targets physical plausibility? This would show whether FCP's self-supervised future targets match or exceed external-encoder alignment.
  6. (Clarification) Results are means over four seeds, but no standard deviations are reported. In Fig. 4, the teacher-forcing curve varies by about 2 IQ points between neighboring checkpoints, which exceeds several ablation gaps (e.g., parallel versus chained predictors, 0.8; with versus without text, 1.1; offset sets in Table 2, 0.7). Could you report standard deviations across seeds for Tables 2–4 and Table 3? This would show which differences are reliable, including the small PhyGround changes.
8 comments
AuthorsOpenReview ↗

Response to W1: Cross-Model Comparisons and Attribution to FCP

We agree that fair cross-model comparisons are challenging, as training data and recipes are often incompletely disclosed even for open-source models, limiting faithful reproduction.

Beyond data differences, our models also undergo adaptation from bidirectional to causal autoregressive generation. As already noted in Section 5.4, we emphasize the controlled comparisons in Tables 2 and 4, while Table 1 provides broader performance context through cross-model benchmarking, following common practice in video generation research. We will add the TF baseline to Table 1 and moderate the corresponding cross-model claims.

We also plan a systematic, controlled study of bidirectional versus causal autoregressive modeling for temporal evolution and will report the findings when resources permit. Our current computational resources are insufficient to complete this study in the short term.

AuthorsOpenReview ↗

Response to Q1: Offset Definitions and Distant-Future Prediction

As the reviewer correctly notes, the offset ranges differ. Eq. 5 conceptually formulates distant-future prediction (Δ>1\Delta>1), whereas Eq. 6 describes our practical implementation (Δ≥1\Delta\ge1). With the full clean history x≤kx_{\le k} as conditioning, Δ=1\Delta=1 recovers standard next-chunk prediction. Conditioning the auxiliary predictor only on hkh_k, however, introduces a distinct auxiliary objective with an additional representation-learning challenge: predicting future chunks from a compact source representation without direct access to the full clean history. This distinction is especially clear in the noisy-source branch, where hkh_k is extracted from a noise-perturbed source chunk rather than its clean counterpart. Thus, although Δ=1\Delta=1 predicts the adjacent future chunk, it still imposes an additional predictive requirement on hkh_k beyond standard TF, motivating its inclusion in Eq. 6.

Another motivation for retaining Δ=1\Delta=1 is our plan to explore lightweight FCP heads for accelerating diffusion sampling, as discussed in Appendix C.2. Our early experiments suggested that adjacent-chunk prediction (Δ=1\Delta=1) achieved the best prediction quality among the tested offsets, consistent with its shorter prediction horizon. While Δ=1\Delta=1 may not be essential for FCP training, we retain it at some additional training cost to support potential inference-time extensions.

In response to the reviewer’s question, we conducted additional offset experiments on the 5B model. The I2V Physics-IQ results are summarized below:

MethodPrediction offsetsPhysics-IQ Verified ↑
Teacher forcing—30.4
FCP — adjacent only{1}30.1
FCP — distant only{2,3,4,5}33.5
FCP — full configuration{1,2,3,4,5}33.8

Adjacent-only FCP (Δ=1\Delta=1) does not improve physical adherence in this setting. By contrast, supervision using only Δ>1\Delta>1 retains most of the full configuration’s improvement over TF, supporting distant-future prediction as the main source of the observed gains. However, our early experiments also showed benefits from offset-1 supervision under some evaluation settings. We therefore still recommend retaining this branch for its potential to support future extensions, depending on the intended application.

AuthorsOpenReview ↗

Response to Q2: Long-Horizon Generation

We agree that reliable long-horizon generation is an important capability for practical applications of autoregressive video models. However, FCP focuses on Stage 1 teacher-forced training, analogous to the first stage of Causal Forcing [1], rather than rollout-based post-training [2]. Since this stage does not explicitly optimize long-horizon rollouts, generation beyond the training horizon may become unstable. Reliable long-horizon generation typically requires additional adaptation, such as dedicated post-training on extended self-generated rollouts [3]. This view is also shared by other researchers in the community [4]. Preliminary results from our ongoing internal post-training study show that FCP retains its physical-consistency advantage over TF under the same post-training protocol. We will prioritize this direction in our next research phase and publicly report the findings.

To address the reviewer’s concern within our available resources, we conducted additional long-horizon experiments in the 1.3B setting, applying the same Helios-inspired [5] history-context noise augmentation during training to both TF and FCP. Over 20-second rollouts, TF and FCP achieve comparable average quality scores of 76.7 and 76.8, respectively, consistent with our focus on modeling future evolution rather than explicitly enhancing visual quality. Since Physics-IQ’s standard five-second evaluation protocol [6] does not directly cover these longer rollouts, we additionally report FVD against real reference videos as a complementary measure of distributional fidelity [7]. During the latter half of the 60-second rollouts (30–60 s), FCP achieves lower FVD than TF in all six five-second segments, reducing mean segment-wise FVD by 8.9% (1088.51 → 991.75). Together, these results support the applicability of FCP to longer rollouts, with comparable video quality and better late-rollout distributional fidelity under the same training adaptation.

References

[1] Zhu et al. Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation. ICML, 2026.

[2] Huang et al. Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion. arXiv:2506.08009, 2025.

[3] Yang et al. LongLive: Real-time Interactive Long Video Generation. arXiv:2509.22622, 2025.

[4] Causal Forcing authors. Response to “Is progressive blurring after ~150 frames expected after the teacher-forcing stage?” Official GitHub repository, Issue #52, 2026.

[5] Yuan et al. Helios: Real Real-Time Long Video Generation Model. arXiv:2603.04379, 2026.

[6] Physics-IQ Benchmark. Official evaluation protocol: Trim Videos. Official GitHub repository.

[7] Unterthiner et al. Towards Accurate Generative Models of Video: A New Metric & Challenges. arXiv:1812.01717, 2018.

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

This paper proposes Future Chunk Prediction (FCP), an auxiliary training objective for autoregressive video diffusion models. Motivated by the temporal redundancy in teacher-forced training, FCP encourages the generator's intermediate representations to support the prediction of temporally distant future chunks without observing the intervening visual context. The approach uses offset-specific auxiliary diffusion predictors that are discarded at inference time.

Experiments on Wan2.1-1.3B and Wan2.2-5B show improvements in physical plausibility over teacher-forced baselines. The paper also investigates architectural choices and provides qualitative evidence of improved spatiotemporal representations.

Critical strengths

Simple and practical approach. The proposed auxiliary objective is straightforward and does not require additional pretrained visual encoders. The auxiliary branches are removed during inference, preserving the original autoregressive generation procedure and inference cost.

Encouraging controlled results. The improvements on Physics-IQ Verified are consistent across model scales and conditioning settings. For example, FCP improves the I2V IQ score from 30.4 to 33.8 on Wan2.2-5B and the V2V score from 40.6 to 44.4.

Useful architectural ablations. The comparisons between clean and noisy source features, parallel and chained predictors, and different conditioning strategies provide valuable insights into the design of future-prediction supervision.

Relevant evaluation focus. Evaluating physical plausibility is well aligned with the stated motivation. The use of reference-based metrics provides useful evidence beyond subjective visual comparisons.

Critical weaknesses
  1. Lack of human evaluation to validate perceptual improvements.

A major concern is the absence of systematic human evaluation. Although FCP achieves consistent improvements on Physics-IQ Verified and several VLM-based benchmarks, it remains unclear whether these numerical gains translate into perceptually meaningful improvements in generated videos.

In particular, the improvements in reference-based physical plausibility metrics do not necessarily imply that human observers would consistently perceive FCP-generated videos as more physically realistic or visually coherent. The qualitative examples in Figure 3 are encouraging but insufficient to establish this claim at scale. Moreover, improvements in physics-related metrics may potentially come at the expense of other aspects of video quality, which are not systematically evaluated.

A human GSB (Good/Same/Bad) evaluation comparing FCP against teacher forcing would be valuable, preferably assessing physical plausibility and overall video quality separately. Without such evaluation, the perceptual significance of the reported improvements remains unclear.

  1. Insufficient ablations to isolate the contribution of distant-future prediction.

The central motivation of FCP is that predicting temporally distant chunks encourages the model to capture visual evolution rather than relying on local temporal redundancy. However, the current experiments do not sufficiently isolate the contribution of distant-future prediction from general auxiliary supervision.

Specifically, the proposed method includes offset 1 in its default configurations, and the improvements could partly arise from additional next-chunk supervision or general multi-task regularization rather than learning longer-range temporal dynamics.

Although the paper examines different numbers and combinations of offsets, it lacks a systematic comparison between next-chunk-only supervision and strictly distant-future supervision under matched computational budgets and auxiliary loss weights. Such experiments would provide more direct evidence supporting the proposed motivation.

  1. Limited methodological novelty and insufficient comparison with closely related work.

The proposed future-chunk prediction objective is closely related to Next Forcing, which also explores future prediction as an auxiliary training signal. While the paper introduces several meaningful design choices, including source-chunk-only conditioning, parallel offset-specific predictors, and supervision of both clean and noisy representations, the fundamental idea of future prediction is shared with this closely related work.

The paper provides useful component-level comparisons against full-context conditioning and chained prediction. However, these experiments do not constitute a complete comparison with Next Forcing under a unified experimental setup.

A more direct comparison would help clarify the independent contribution of FCP and determine how much of the reported improvement comes from the proposed architectural choices rather than the shared future-prediction training principle.

  1. Limited evaluation of long-term temporal consistency and generalization.

The paper motivates FCP as a means of improving the modeling of temporal evolution in autoregressive video generation. However, the main evaluations focus on five-second video continuations, leaving its effectiveness over longer autoregressive rollouts largely unexplored.

Long-horizon generation is particularly important in this setting, as autoregressive models may suffer from accumulated errors, temporal inconsistencies, and progressively degraded physical plausibility. It remains unclear whether FCP can mitigate these problems or whether its improvements are primarily limited to short-term generation.

Furthermore, although FCP generally improves the reported results on complementary benchmarks, some gains are relatively small, such as +0.20 on PAI-Bench-G Physics and +0.05 on PhyGround Physics.

Evaluations over longer generation horizons, together with an analysis of how temporal consistency and physical plausibility evolve over time, would strengthen the paper's broader claims regarding visual evolution and world modeling.

  1. Incomplete computational and statistical analysis.

The paper emphasizes the training efficiency of FCP and reports substantial estimated FLOPs savings compared with the full-context variant. However, the proposed approach introduces multiple independently parameterized auxiliary diffusion branches during training, potentially increasing training-time memory consumption and implementation complexity.

While the estimated-FLOPs comparison is informative, the practical computational overhead is not fully characterized. Reporting actual GPU-hours, peak memory usage, and a detailed computational breakdown would help establish the practical efficiency of the proposed method.

Additionally, although the authors evaluate generated videos using four random seeds, the reported results primarily consist of mean scores without uncertainty estimates. The absence of standard deviations or confidence intervals makes it difficult to assess the reliability of smaller improvements, especially on the complementary benchmarks.

More comprehensive computational reporting and statistical analysis would make the empirical conclusions more convincing.

Questions for authors

Q1. Human Evaluation. Have the authors conducted any human preference evaluations to verify whether the improvements on physics-related benchmarks translate into perceptually meaningful gains? In particular, could the authors provide a pairwise GSB (Good/Same/Bad) comparison between FCP and the teacher-forcing baseline, separately evaluating physical plausibility and overall video quality? This would help establish whether the reported numerical improvements are consistently noticeable to human observers.

Q2. Contribution of Distant-Future Prediction. Since the central motivation of FCP is to encourage modeling of long-range temporal evolution, how much of the improvement can be attributed specifically to predicting distant future chunks? Could the authors compare supervision using only offset 1, only offsets greater than 1, and the full configuration, while controlling for auxiliary loss weights and computational budgets? Such an ablation would help distinguish the benefits of distant-future prediction from those of general auxiliary supervision.

Q3. Comparison with Next Forcing. While the paper discusses the differences from Next Forcing and provides component-level ablations, could the authors clarify how FCP compares with the complete Next Forcing approach under matched training and evaluation settings? In particular, how much of the improvement can be attributed to the proposed parallel prediction and source-chunk-only conditioning strategies?

Q4. Long-Horizon Generation. Does FCP maintain its advantages over teacher forcing when generating substantially longer videos? Given the motivation of improving temporal evolution modeling in autoregressive generation, it would be valuable to evaluate whether the improvements in physical plausibility and temporal consistency persist over longer rollouts, where accumulated errors become more significant.

Q5. Computational Cost and Statistical Reliability. Could the authors provide additional details on the practical training overhead of FCP, including GPU-hours and peak memory consumption compared with teacher forcing? Additionally, since some improvements on complementary benchmarks are relatively small, could the authors report standard deviations or confidence intervals across the evaluation seeds to better assess their statistical reliability?

7 comments
AuthorsOpenReview ↗

Response to Q1: Human Evaluation and Video Quality

We appreciate the reviewer’s concern about perceptual relevance. Physics-IQ Verified scores closely align with human judgments at the model level, as demonstrated by an independent Rapidata study [1]. WorldModelBench, PAI-Bench-G, and PhyGround further support our evaluation framework through comparisons of their automatic assessments with human annotations or preferences [2] [3] [4].

These prior studies provide empirical support for our evaluation using established benchmarks. At the same time, our current staffing and resource constraints prevent us from conducting a dedicated human study with independent evaluators, standardized training, and sufficient quality control. We acknowledge this limitation and plan to incorporate systematic human evaluation as our method develops toward more mature applications, assessing physical plausibility and overall video quality separately.

Regarding video quality, we agree that it should not be sacrificed for improvements in physical adherence. To address this concern, we evaluated video quality using the VBench-based quality metrics in PAI-Bench-G [3]. The overall score is 76.77 for teacher forcing and 76.78 for FCP, indicating essentially tied overall video quality. Thus, FCP improves physical adherence while preserving comparable generation quality, consistent with our focus on modeling future evolution without introducing a separate objective specifically aimed at enhancing visual quality.

Quality metricTeacher forcingFCP
Aesthetic quality50.0350.17
Background consistency93.9493.66
Imaging quality65.1064.93
Motion smoothness99.0298.96
Overall consistency21.2421.40
Subject consistency91.6791.55
I2V background consistency97.0597.23
I2V subject consistency96.0896.30
Overall score76.7776.78

References

[1] Rapidata. World Models: Physics Generation Benchmark v2. Online research report, 2026.

[2] Dacheng Li et al. WorldModelBench: Judging Video Generation Models As World Models. NeurIPS, 2025.

[3] Fengzhe Zhou et al. PAI-Bench: A Comprehensive Benchmark For Physical AI. arXiv:2512.01989, 2025.

[4] Juyi Lin et al. PhyGround: Benchmarking Physical Reasoning in Generative World Models. arXiv:2605.10806, 2026.

AuthorsOpenReview ↗

Response to Q2: Contribution of Distant-Future Prediction

We conducted additional experiments with different prediction-offset configurations on the 5B model. The I2V Physics-IQ results are summarized below:

MethodPrediction offsetsPhysics-IQ Verified ↑
Teacher forcing—30.4
FCP — adjacent only{1}30.1
FCP — distant only{2,3,4,5}33.5
FCP — full configuration{1,2,3,4,5}33.8

Adjacent-only FCP (Δ=1\Delta=1) does not improve physical adherence in this setting. We retain this branch for its potential to accelerate inference and the benefits observed in some preliminary experiments and other settings. Its representation-conditioned objective also differs from standard next-chunk prediction, as explained in our response to Reviewer PWCw (Q1).

By contrast, supervision using only Δ>1\Delta>1 retains most of the full configuration’s improvement over TF, supporting distant-future prediction as the main source of the observed gains. These findings provide strong empirical support for our motivation to improve autoregressive video generation by explicitly constraining more distant future evolution.

We further compared offset configurations under matched computation, using three auxiliary predictors in each case:

Prediction offsetsPhysics-IQ Verified ↑
{1,2,3}32.1
{1,3,5}33.1
{1,4,7}33.2

These results show that extending the prediction horizon can yield additional gains at the same computational budget. Although we observed no decline within the tested range, we expect gains to eventually saturate or reverse, as predicting increasingly distant futures without access to intervening context becomes increasingly underdetermined. Our exploration was limited by computational resources during the rebuttal and the scarcity of valid supervision pairs at larger offsets in 5 s training clips. For example, at Δ=9\Delta=9, only the first chunk of each sample can serve as a prediction source, making stable training more difficult.

An exhaustive search for the optimal efficiency–performance trade-off is beyond our current resources. Our experience suggests that this balance depends on the model, task, and training data, making the appropriate configuration application-dependent. As an initial study, we aim to provide a broadly applicable design and practical configurations that enable practitioners to obtain consistent improvements across diverse settings.

AuthorsOpenReview ↗

Response to Q3: Comparison with Next Forcing

Our manuscript already examines key design choices relevant to Next Forcing through controlled ablations: chained versus parallel prediction (Tables 2), full-history versus source-chunk-only conditioning (Fig. 4), and explicit text conditioning in the auxiliary predictors (Table 4).

To complement these component ablations with the requested full-method comparison, we faithfully reproduced the complete Next Forcing training approach during the rebuttal, adapting it to our autoregressive video-generation setting. The results on 1.3B models are:

MethodPhysics-IQ ↑
Teacher forcing27.6
Next Forcing28.8
FCP30.9

Both future-prediction approaches improve over TF, while FCP achieves a further 2.1-point gain over Next Forcing under the same training setup. Together with the component ablations, these results support the effectiveness of FCP’s specific design beyond the shared future-prediction principle.

Our work was developed independently and concurrently with Next Forcing, with the initial FCP prototype completed before Next Forcing became publicly available. While Next Forcing primarily focuses on world-action modeling, FCP targets autoregressive video generation, motivating different design choices and evaluation objectives. We hope this independent and concurrent development can be considered in light of ICLR’s guidance on contemporaneous work [1]. We also provide the direct comparison above to help clarify the relationship between the two approaches and FCP’s contribution.

References

[1] ICLR 2027. Reviewer Guidelines, FAQ on contemporaneous and non-peer-reviewed work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank You for Your Time and Thoughtful Feedback

Dear Reviewers,

Thank you for the time and care you have devoted to our paper. As the discussion draws to a close, we would appreciate hearing whether our additional experiments and explanations have addressed your main concerns. We would be happy to clarify any remaining questions.

Thank you again for your constructive suggestions. They have helped us strengthen both our experiments and our presentation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.