acceptodds
Withdrawn from ICLR 2027

HorizonDrive: Self-Corrective Autoregressive World Model for Long-horizon Driving Simulation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Closed-loop driving simulation requires real-time interaction beyond short offline clips, pushing current driving world models toward autoregressive (AR) rollout. As generated frames become the context for subsequent chunks, prediction errors accumulate over time. The key difficulty is that a teacher trained only on clean histories provides unreliable corrective guidance when conditioned on the student's generated frames. A natural question is: can a teacher provide reliable corrective guidance throughout a student's long AR rollout? We address this with HorizonDrive, a teacher-centric training-and-distillation framework for AR driving simulation. Through scheduled rollout recovery (SRR), we mine semantically degraded yet layout-preserving states from long driving rollouts, turning their retained scene structure into corrective supervision for training a recovery-capable teacher. Further, teacher rollout distillation (TRD) transfers this capability to a short-chunk, few-step student through fixed-window distribution matching, enabling continual correction throughout long AR rollouts with bounded memory and per-window computation. HorizonDrive natively supports minute-scale AR rollout under bounded memory; on nuScenes, HorizonDrive reduces FID by 52% and FVD by 37%, and lowers ARE and DTW by 21% and 9% relative to the strongest long-horizon streaming baselines. Open- and closed-loop evaluations further show improved controllability and downstream planning utility.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, Close-loop Simulation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_JnAeOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a self-corrective autoregressive world model for efficient long-horizon driving video generation. The method introduces Scheduled Rollout Recovery (SRR) to improve the teacher model's robustness against accumulated autoregressive errors, and Teacher Rollout DMD (TRD) to distill long-horizon generation capabilities into an efficient student model. Experiments on nuScenes and an in-house driving dataset demonstrate improvements in video generation quality and temporal consistency over existing streaming video generation methods.

Critical strengths

The paper addresses error accumulation and temporal inconsistency in long-horizon autoregressive world models, which are important challenges for autonomous driving simulation. The teacher-side recovery strategy provides an interesting perspective on stabilizing autoregressive generation before distillation. The method achieves competitive FID/FVD results against existing approaches, with relatively fair comparisons using a shared backbone and supporting ablation studies.

Critical weaknesses
  1. Limited Novelty: The proposed SRR and TRD mainly extend existing self-forcing, recovery training, and diffusion distillation techniques. The individual contributions of the SRR components are not sufficiently validated.
  2. Insufficient Long-horizon Evaluation: Quantitative evaluation mainly covers approximately 20-second sequences, while minute-level demonstrations are largely qualitative. Long-term error accumulation and physical consistency remain insufficiently investigated.
Questions for authors

See the "Weaknesses"

4 comments
AuthorsOpenReview ↗

Response to Reviewer JnAe (Part 1/2)

We thank the reviewer for recognizing the importance of error accumulation and the teacher-side recovery perspective. We clarify the methodological contribution and point to the component-level and longer-horizon evidence below.

[W1] Novelty and the contributions of the SRR components

a. Differences from existing long-video frameworks

(1) Compared with forcing methods such as Self-Forcing++ and LongLive, the distinction is whether the teacher also undergoes targeted recovery training. These methods primarily optimize long-horizon generation by the few-step student. HorizonDrive first uses SRR to strengthen the teacher's correction capability on degraded histories and then distills the student, rather than requiring only the student to adapt to accumulated errors without improving the teacher's guidance under those histories.

(2) Compared with recovery training methods such as SVI, SRR differs in its error source and intended application. SVI resamples and injects errors from an error bank. SRR instead uses degraded histories that actually accumulate during the teacher's continuous long-horizon rollouts, training the teacher to recover from semantic errors while preserving trajectory continuity, and then using the enhanced teacher for few-step student distillation.

HorizonDrive therefore contributes two successive technical designs: SRR strengthens teacher recovery on degraded histories while preserving semantic fidelity and trajectory continuity; TRD combines the enhanced teacher with CFG scheduling to reliably transfer its long-horizon capabilities to a few-step student.

The systematic evaluation further validates the practical value of this design. HorizonDrive achieves FID/FVD/ARE/DTW of 13.82/92.99/2.60/3.27, outperforming both Self-Forcing++ (28.84/147.57/3.78/3.61) and LongLive (29.05/161.41/3.28/3.65) across all four metrics. We believe these methods and experiments together reveal an important optimization direction that prior work has underexplored: providing the student with a teacher capable of long-horizon generation.

b. SRR's internal design and comparison with existing recovery methods

Are both curricula in SRR important? Table 3 further validates the two curricula. Under the same uniform boundary sampling, replacing the fixed transition radius w=4w=4 with a progressively increasing radius w:0→8w:0\to8 reduces FID/FVD from 16.20/132.51 to 15.52/103.02, while also improving the geometric metrics. Adding boundary sampling decay further reduces FID/FVD to 13.73/94.84 and ARE/DTW from 2.05/3.16 to 1.95/2.99, showing that both curricula provide additional benefits.

Does SRR offer advantages over existing recovery training methods? Table 8 directly compares SRR with Stable Video Infinity (SVI) [1] under the same teacher architecture, initialization, data, and training budget. SRR achieves FID/FVD of 13.73/94.84, compared with SVI's 25.39/173.98, and reduces ARE/DTW from 2.61/3.31 to 1.95/2.99. These results show that, under this driving-generation setting, SRR achieves better visual quality and geometric consistency than the existing recovery training method.

c. SRR's contribution to student distillation and long-horizon stability

Table 4 fixes the TRD settings and student initialization. Replacing only the Base teacher with the SRR teacher reduces FID/FVD from 19.24/141.88 to 14.44/107.54; initializing both the teacher and student from SRR further reduces them to 13.82/92.99. These results demonstrate SRR's contribution to the quality of student distillation.

The minute-scale chunk-FID curves in Figure 6 further show that full HorizonDrive exhibits slower visual-quality degradation than the variant without SRR and remains relatively stable beyond the training horizon, supporting SRR's contribution to stable long-horizon generation.

AuthorsOpenReview ↗

Response to Reviewer JnAe (Part 2/2)

[W2] Long-horizon evaluation and consistency

a. Additional one-minute quantitative evaluation

We additionally evaluated the existing models, trained with 20-second rollouts, on 100 one-minute videos from our internal dataset:

MethodFID ↓FVD ↓ARE ↓DTW ↓
Self-Forcing189.481534.208.4817.01
Self-Forcing++200.401369.7312.0318.43
LongLive96.64792.296.429.97
HorizonDrive19.21188.833.063.73

In the one-minute generation setting, HorizonDrive outperforms the compared methods on all four metrics: FID, FVD, ARE, and DTW. These results extend minute-scale validation from qualitative demonstrations to quantitative evaluation, supporting the visual quality of long-horizon generation and providing additional evidence of spatial consistency through lower ARE and DTW.

b. Chunk-wise evaluation: does error accumulate over generation time?

To illustrate long-horizon error accumulation, Figure 6 shows how chunk FID changes with generated video length. The values are:

Method10 s20 s30 s40 s50 s60 s
LongLive30.8945.7358.5164.4569.3973.33
HorizonDrive w/o SRR20.7224.9128.9831.4933.3234.99
HorizonDrive18.1117.4018.0019.3319.4519.88

HorizonDrive maintains low and stable chunk FID throughout the one-minute rollout, whereas the chunk FID of LongLive and HorizonDrive w/o SRR increases over time. These results support that, through the combined optimization of SRR and TRD, HorizonDrive maintains stable visual quality beyond the 211-frame training horizon.

References

[1] W. Li et al. Stable Video Infinity: Infinite-Length Video Generation with Error Recycling. ICLR, 2026.

AuthorsOpenReview ↗

Follow-up Response to Reviewer JnAe

Thank you again for your helpful feedback. We wanted to follow up on our response and see if you have any remaining questions or concerns. We’d be happy to clarify any points.

AuthorsOpenReview ↗

Follow-up on Novelty and SRR Validation

As the discussion deadline approaches, we would be grateful for your thoughts on the technical comparisons, SRR component ablations, and one-minute quantitative results presented in our rebuttal. If concerns remain about novelty or component validation, it would be helpful to know which specific aspects require further clarification so that we can address them. Thank you for your time and consideration.

Reviewer_4LoUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present HorizonDrive, a driving video generator built on Wan 2.1 that takes an HD map, 3D boxes and the ego action and generates video chunk by chunk. The core idea is to handle exposure bias on the teacher side. The base model is rolled out on its own predictions, the degraded histories are cached, and the model is fine-tuned to predict the logged future from them (SRR). A training schedule controls how this switch from predicted to real frames is done. This teacher is then distilled into a four-step student with DMD on the student’s own rollouts, with a modified CFG schedule (TRD). The claim is that a recovery-trained teacher gives reliable guidance on drifted states, which yields stable, controllable rollouts at minute scale with bounded memory, usable for closed-loop simulation. On nuScenes the authors report large FID/FVD/ARE/DTW gains over retrained Self-Forcing, Self-Forcing++ and LongLive, plus perception and planner metrics, a 10-step planner-in-the-loop test, a second internal dataset and qualitative one-minute rollouts. The student runs at 5.6 FPS at 256×512 and 1.7 FPS at 384×768, against 10 FPS data.

Critical strengths

S1. The idea of making the teacher robust to its own rollout errors before distilling is a sensible one, and the paper actually tests it properly. Table 4 swaps teacher and student initialization independently: using the SRR teacher alone brings FID/FVD from 19.2/141.9 down to 14.4/107.5, and using it for both gets to 13.8/93.0. That is a clean result and the part of the paper I find most convincing. The comparison against a reimplemented SVI under the same budget (Table 8) is also appreciated. S2. The evaluation is broader than I expected: per-chunk quality over time rather than only an aggregate number, a second dataset, downstream perception and planner metrics, and throughput numbers on a single consumer GPU.

Critical weaknesses

Weaknesses W1a. Attribution. Table 4 row 1 is TRD without SRR and it already beats all baselines in Table 1 on all four metrics. So most of the FID/ARE/DTW gain over the baselines is from the distillation setup, not SRR. SRR itself gives the FVD gain and about a third of the FID gain, ARE/DTW barely move. The paper still presents SRR as the main contribution. This is not a reason to reject, but the claims should match Table 4.

W1b. I am not sure the baseline comparison is fair. The baselines run with their original recipes (Sec. F), while HorizonDrive’s distillation settings are tuned. If I read Sec. 3.3 correctly, restricting CFG to low noise levels is this paper’s own change (following Decoupled DMD), so the baselines presumably use it at all noise levels. Table 7 shows that with full-range CFG the authors’ own model drops to FVD 184 and ARE 3.86, i.e. baseline level or worse. So at least part of the difference to the baselines may come from a distillation setting rather than from SRR. There are also no seeds anywhere, and the Table 4 ablation rows move by more than some of the headline differences (the DTW gain over SF++ is 0.34), so it is hard to tell which of the smaller differences are real.

W2. The long-horizon and simulation claims go beyond what is shown. The student is trained with 20-chunk rollouts and evaluated on 20-chunk rollouts, so there is no number past the training horizon. “Minute-scale” is two qualitative examples on a private dataset. I also think the paper underplays how much the conditioning does here. Giving the map and boxes for future frames at every step is a reasonable design, but it means the layout never has to be predicted, and every metric is computed with the logged actions and layouts. Nothing tests what happens when the ego does something the log didn’t, except the 10-step planner test, where the planner more or less follows the log anyway. Related to this, as far as I understand Sec. E, ARE and DTW are computed from camera poses recovered from the generated video while the ego action is an input, so they are a good controllability check but say less about whether the rest of the scene stays consistent. Lastly, Epona is cited and even used as the planner, reports FID 7.5 / FVD 82.8 on nuScenes with rollouts of two minutes, and does not appear in Table 5 (which already mixes protocols), which seems at odds with Fig. 1’s claim that driving world models are “limited in AR rollout”.

W3. I don’t know what to make of the planner results (Tables 12, 13). “Closed-loop” here means the planner’s trajectory is fed back as the ego action, the other agents and the map are still replayed from the log, and PDMS itself is non-reactive. Fig. 15 says the loop is “driven entirely by self-generated signals”, which is not quite accurate, since we are still having log-replay. PDMS is a NAVSIM metric and the paper does not say how it was computed on nuScenes (drivable area, TTC, map, horizon) or on how many scenes. On top of that, nuScenes open-loop planning is known to be solvable from ego status alone (AD-MLP, Zhai et al. 2023; Li et al. CVPR 2024), so this experiment’s signal is not strong. The expert row makes me wonder about the implementation: the logged human trajectory gets a 0.31% collision rate against its own replayed agents, so the check has false positives about as large as the differences between methods.

Questions for authors

Q1 (decides the score). Could the baselines be run with the same distillation settings as HorizonDrive (the CFG schedule and the training rollout length), so the comparison isolates SRR? Adding TRD with the base teacher and student to Table 1 would help for the same reason. Results over a few seeds would also help to see which differences are real. Q2 (decides the score). Could you report per-chunk quality beyond 20 chunks on the e2e data, and add the TRD-without-SRR model to Fig. 6? That would directly show whether SRR reduces drift over time. Q3 (decides the claims). Could you show what happens under actions that differ from the log, using the Sec. N setup (e.g. slower, faster, hard brake, lateral offset)? Some number on how well the generated ego motion follows the commanded trajectory, plus a few videos, would be enough. If that’s not possible, I’d suggest describing the setting as “ego closed-loop, non-reactive” instead of closed-loop simulation.

Clarifications How PDMS is computed on nuScenes and on how many scenes. Epona’s score on real frames and on blank or shuffled frames, to show the score depends on the images and Epona’s planning at all. Table 10 lists 200 training steps but a CFG decay that ends at step 400. Which one is right? As far as I can tell from the MagicDrive-V2 paper, its FVD 94.84 is computed on the first 16 frames and its FID is from the image evaluation, so neither is a 241-frame number. Could you check this? Sec. 4.2’s comparison rests on it.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 4LoU (Part 1/2)

We thank the reviewer for the feedback; we have clarified SRR’s contribution, strengthened baseline comparisons and long-horizon evaluation, and corrected and expanded downstream experiments to address concerns about simulation utility.

[W1/Q1/Q2] Metric clarification and contribution attribution

a. Why does Base/Base already outperform the baselines?

Base/Base shares the same training rollout length and teacher initialization as Self-Forcing++ and LongLive. The main differences are that TRD uses a bidirectional rather than causal student, as in BiWM [1], and introduces CFG threshold decay.

b. Why is SRR + full-range CFG comparable to the baselines? Does this weaken SRR's contribution?

Both SRR and CFG optimization within TRD are key contributions of HorizonDrive: the former strengthens the teacher's recovery capability on degraded histories, while the latter adjusts how guidance is used during distillation. Enhancing the teacher with SRR alone may therefore still yield distribution shifts during distillation, motivating our joint design.

SRR's contribution should be assessed with and without SRR under identical CFG settings. Table 4 fixes the TRD settings and student initialization: replacing only the Base teacher with the SRR teacher reduces FID/FVD from 19.24/141.88 to 14.44/107.54, directly demonstrating the additional benefit of teacher recovery training.

Nevertheless, SRR remains important for error robustness. Figure 6 compares the chunk-FID curves of Base/Base without SRR and full HorizonDrive. Full HorizonDrive maintains stable long-horizon quality, whereas CFG optimization alone still suffers from error accumulation.

c. How does the baseline with the same CFG optimization compare with HorizonDrive?

We retrained LongLive with CFG threshold decay and compared it with the original LongLive and HorizonDrive:

MethodFID ↓FVD ↓ARE ↓DTW ↓
LongLive29.05161.413.283.65
LongLive + CFG decay22.80163.932.894.58
HorizonDrive13.8292.992.603.27

CFG decay improves LongLive's FID and ARE, but performance remains limited by the teacher's capabilities and still falls short of full HorizonDrive.

d. Comparison across multiple seeds

We added evaluations with inference seeds 0 and 1. HorizonDrive shows clear advantages across all metrics and retains its ranking across both seeds.

MethodFID ↓FVD ↓ARE ↓DTW ↓
Self-Forcing41.11 ± 0.51167.60 ± 1.553.38 ± 0.085.72 ± 0.33
Self-Forcing++25.08 ± 0.61150.14 ± 2.763.15 ± 0.063.78 ± 0.15
LongLive26.62 ± 0.41175.98 ± 0.382.65 ± 0.013.82 ± 0.09
HorizonDrive13.81 ± 0.00185.38 ± 2.682.17 ± 0.053.10 ± 0.18

[W2/Q2/Q3] Long-horizon evaluation, baseline comparisons, and control conditions

a. Missing quantitative evaluation of minute-scale generation

We evaluated the existing models, trained with 20-second rollouts, on 100 one-minute videos from our internal dataset:

MethodFID ↓FVD ↓ARE ↓DTW ↓
Self-Forcing189.481534.208.4817.01
Self-Forcing++200.401369.7312.0318.43
LongLive96.64792.296.429.97
HorizonDrive19.21188.833.063.73

Figure 6 plots chunk FID against video length to illustrate error accumulation. HorizonDrive maintains low and stable chunk FID throughout the one-minute rollout, whereas LongLive and HorizonDrive w/o SRR show increasing chunk FID over time. These results support HorizonDrive's stable visual quality beyond the 211-frame training horizon:

Method10 s20 s30 s40 s50 s60 s
LongLive30.8945.7358.5164.4569.3973.33
HorizonDrive w/o SRR20.7224.9128.9831.4933.3234.99
HorizonDrive18.1117.4018.0019.3319.4519.88

b. The role of control conditions and testing beyond logged actions

We acknowledge the role of control conditions and provide all baselines with the same controls for a fair comparison. We study ego-action-controlled visual rollouts under externally provided layouts, rather than asking the model to independently predict all traffic participants' future behavior.

The appendix and supplementary material show results under counterfactual ego actions. Figure 15 shows a closed-loop simulation deviating from the original viewpoint, with several similar supplementary videos. These examples show that HorizonDrive maintains visual stability even when deviating from the log. We also added Figure 16, comparing observations generated from the same starting point under four actions: acceleration, deceleration, hard braking, and a lane change. Corresponding videos are included in the supplementary material.

AuthorsOpenReview ↗

Response to Reviewer 4LoU (Part 2/2)

[W2/Q2/Q3] Long-horizon evaluation, baseline comparisons, and control conditions (continued)

c. Controllability and scene consistency

HorizonDrive is conditioned on ego actions, HD maps, and bounding boxes. Table 2 complements ARE/DTW by evaluating road, lane, and object information, as well as planner responses. Road/lane IoU and box mAP compare YOLOP and FCOS3D outputs on generated images with those on corresponding real images; ADE/FDE compare Epona trajectories predicted from the two observation sources. HorizonDrive achieves the best results across all five metrics, supporting its scene consistency.

d. Epona's long-video evaluation and the characterization of prior models' AR capabilities

We agree that Epona's results should be included and that Figure 1 characterizes prior driving models' AR capabilities too broadly.

We re-evaluated Epona on 150 paired nuScenes videos, conditioned on the first 10 images and subsequent actions. At an evaluation length of 208 frames, it achieves FID/FVD of 31.94/167.68, compared with HorizonDrive's 13.82/92.99 over 211 frames. These results are included in Table 5 of the revised manuscript. We will also revise our description and positioning of existing methods.

[W3/C1/C2] Corrections to downstream evaluation and the planner's dependence on visual inputs

We thank the reviewer for identifying issues in our downstream evaluation. Code inspection revealed inconsistent BEV coordinate mappings between trajectory points and obstacle occupancy grids, causing incorrect collision detection. We corrected the mapping, expanded the set of evaluation starting points, and recomputed the affected metrics. Implementation details and the evaluation protocol are provided in Appendix N.

a. Corrected open- and closed-loop evaluation

Evaluation protocol: Open-loop evaluation involves no environment feedback and scores the planner's 1.5-second trajectory predictions at 10 anchors in each of 150 scenes. Closed-loop evaluation scores the complete paths actually executed by the planner, with ego actions providing feedback while other traffic participants follow their logged trajectories.

Open-loop results: The corrected results are shown below. Expert GT's collision rate drops to zero. Among the compared generation methods, HorizonDrive still yields the best planner results.

Method / observation inputCollision rate (%) ↓EP ↑PDMS ↑Trajectory L2 (m) ↓
Self-Forcing0.140.8379.151.23
Self-Forcing++0.080.8580.421.05
LongLive0.090.8681.040.91
HorizonDrive0.060.8681.410.88
Expert GT0.000.9084.710.00

Closed-loop results: The corrected results below have been incorporated into Table 14 of the revised manuscript.

MethodFrame collision ↓Ego progress ↑Comfort ↑PDMS ↑
Self-Forcing1.22%0.710.8273.81
Self-Forcing++0.82%0.760.8476.13
LongLive1.00%0.780.8678.22
HorizonDrive0.79%0.810.8683.41

b. Does Epona actually depend on visual inputs?

Despite having access to historical trajectories, Epona is affected by visual inputs: blanking all images or changing their order degrades the final evaluation performance.

Visual inputL2 (m) ↓
HorizonDrive0.8802
Real images0.8502
All ten frames blank4.0640
Temporally shuffled real images1.0004

[C3] Training steps and CFG threshold decay

The reported checkpoints for LongLive, Self-Forcing++, and HorizonDrive are taken at 200 training steps, whereas 400 is the step at which the CFG threshold reaches its minimum. In our experiments, the selected checkpoints can precede the completion of the decay schedule.

[C4] MagicDrive-V2 evaluation

We thank the reviewer for pointing out this oversight. MagicDrive-V2's originally reported metrics indeed use a different evaluation protocol from what we previously assumed. We re-evaluated its 241-frame videos on the full nuScenes validation set, obtaining FID/FVD of 24.25/128.96. These results are included in Table 5 of the revised manuscript.

References

[1] S. Rui et al. BiWM: Advancing Open-Source Interactive Video World Models with Bidirectional Autoregression. arXiv preprint arXiv:2606.10135, 2026.

AuthorsOpenReview ↗

Follow-up Response to Reviewer 4LoU

Thank you again for your helpful feedback. We wanted to follow up on our response and see if you have any remaining questions or concerns. We’d be happy to clarify any points.

AuthorsOpenReview ↗

Follow-up on Baseline Fairness and Long-Horizon Evaluation

As the discussion deadline approaches, we would appreciate your feedback on the additional results addressing Q1 and Q2: LongLive retrained with the same CFG schedule, and one-minute quantitative evaluations including the variant without SRR. Could you let us know whether these results address your main concerns? We would be glad to provide any further clarification. Thank you for your time and consideration.

Reviewer_32KvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HorizonDrive tackles error accumulation in autoregressive driving video generation. The core idea is to make the teacher robust first. SRR fine-tunes a controllable driving world model (based on Wan 2.1 1.3B) on its own rollout histories paired with ground-truth futures. It uses two curricula: one over the blending radius at the history–target boundary, and one over the rollout position of that boundary. TRD then distills this teacher into a four-step student with DMD along the student's own 20-chunk rollout, plus a noise-truncated CFG schedule. Experiments cover nuScenes and an internal dataset, with gains in FID/FVD and pose consistency, and some downstream planner evaluations.lity and geometric consistency, along with some planning results and minute-long qualitative videos.

Critical strengths
  1. Well-controlled comparison with large gains. Self-Forcing, Self-Forcing++, LongLive, and SVI are all retrained with the same backbone, data, and control modules. Under this setting, HorizonDrive reduces FID from 28.84 to 13.82 and FVD from 147.57 to 92.99, and the per-chunk FID curve (Fig. 6) shows the improvement comes mainly from late-horizon stability, which is exactly what the method targets.
  2. A sensible, underexplored angle. Existing methods distill from a teacher trained only on clean histories, whose guidance becomes unreliable on degraded student rollouts. Making the teacher itself recovery-capable before distillation is a reasonable fix to this gap.
Critical weaknesses
  1. The gain is not clearly attributable to SRR. In Tab. 4, TRD with a base teacher and base student already reaches 19.24 FID / 141.88 FVD / 2.76 ARE / 3.30 DTW, beating every streaming baseline in Tab. 1 on all four metrics. So roughly 9.6 of the 15-point FID improvement comes from the TRD setup (40-frame teacher window, CFG schedule, N=20) rather than the recovery-trained teacher. Since SRR is presented as the core contribution, the paper should explain how TRD with a base teacher differs from Self-Forcing++/LongLive and which component accounts for this gap.
  2. Long-horizon and closed-loop claims lack quantitative support. All quantitative results cover ~20 s (211 frames), which equals the training rollout length, so there is no evidence of generalization beyond the training horizon. Minute-scale generation is only shown qualitatively in two internal scenes. The closed-loop evaluation runs only 10 steps, compares only against Self-Forcing++ with an identical collision rate (4.58%), and it is unclear whether other agents are log-replayed once the ego deviates. Quantitative results beyond 20 chunks and a planner-ranking consistency test would be needed to support the title's claim.
Questions for authors
  1. In Tab. 4, TRD with a base teacher already beats all the streaming baselines in Tab. 1. What is different from Self-Forcing++ / LongLive in this setting, and did the baselines get a similar training budget? This decides how much of the gain I can credit to SRR.
  2. Do you have any numbers beyond 20 chunks, e.g. per-chunk FID on the internal dataset? Right now the eval horizon is the same as the training rollout length, so I can't tell whether the model actually generalizes to longer horizons.
  3. In the closed-loop test, once the ego leaves the logged trajectory, where do the other agents' boxes come from? If they are just log replay, the "interactive simulation" claim should be toned down.
  4. Tab. 10 says TRD runs for 200 steps, but the CFG threshold decays until step 400. Is one of these a typo?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 32Kv

We appreciate the reviewer’s thoughtful comments on attribution and evaluation; our response clarifies the training differences and SRR’s contribution, provides one-minute results and planner-ranking evidence, and explains the interaction protocol and CFG schedule.

[W1/Q1] Clarifying Baseline Settings and SRR’s Contribution

Base/Base shares the same training rollout length and teacher initialization as Self-Forcing++ and LongLive. The main differences are that TRD uses a bidirectional rather than causal student, as in BiWM [1], and introduces CFG threshold decay. LongLive, Self-Forcing++, and HorizonDrive are trained for 200 steps, while Self-Forcing is trained for 1,000 steps to match the cumulative number of DMD backward passes.

a. Does SRR provide additional benefits with the TRD settings fixed?

Table 4 fixes the TRD settings and student initialization, varying only whether the teacher undergoes SRR:

Student initializationTeacher initializationFID ↓FVD ↓
BaseBase19.24141.88
BaseSRR14.44107.54

Enhancing only the teacher reduces FID/FVD from 19.24/141.88 to 14.44/107.54, directly supporting the additional contribution of teacher-side SRR under identical distillation settings.

b. What happens when the baseline uses the same CFG schedule as TRD?

We additionally retrained LongLive with the same CFG threshold decay:

MethodFID ↓FVD ↓ARE ↓DTW ↓
LongLive29.05161.413.283.65
LongLive + CFG decay22.80163.932.894.58
HorizonDrive13.8292.992.603.27

CFG decay improves LongLive's FID and ARE, but a clear gap to full HorizonDrive remains.

[W2/Q2] Evaluation beyond the training horizon

We evaluated the existing models, trained with 20-second rollouts, on 100 one-minute videos from our internal dataset:

MethodFID ↓FVD ↓ARE ↓DTW ↓
Self-Forcing189.47581534.20108.480117.0080
Self-Forcing++200.39801369.730212.028318.4265
LongLive96.6420792.28576.42129.9652
HorizonDrive19.2105188.82573.05963.7321

To illustrate error accumulation, Figure 6 shows how chunk FID changes with the generated video length. HorizonDrive maintains low and stable chunk FID throughout the one-minute rollout, whereas the chunk FID of LongLive and HorizonDrive w/o SRR increases over time. These results support that, through the combined optimization of SRR and TRD, HorizonDrive maintains stable visual quality beyond the 211-frame training horizon. The chunk-FID values at different rollout times are reported below:

Method10 s20 s30 s40 s50 s60 s
LongLive30.8945.7358.5164.4569.3973.33
HorizonDrive w/o SRR20.7224.9128.9831.4933.3234.99
HorizonDrive18.1117.4018.0019.3319.4519.88

[W2] Planner-ranking consistency

Using observations generated by HorizonDrive, we compared Epona and DriveVA [2]:

PlannerPDMS ↑L2 ↓
Epona81.410.88
DriveVA83.250.44

DriveVA performs better across all metrics, consistent with the results reported in Table 3 of its paper. This provides additional support for the feasibility of using HorizonDrive as a simulation environment.

[W2/Q3] How other agents respond during closed-loop evaluation

In our current setting, other traffic participants follow their logged trajectories, and changes in the ego vehicle's actions do not affect their behavior. We will therefore explicitly describe this setting as “ego closed-loop with non-reactive traffic participants” in the revised manuscript.

[Q4] Training steps and the CFG threshold decay schedule

These two values refer to different settings: 200 steps denotes the checkpoint used for the reported evaluation, while 400 steps is the predefined endpoint of the CFG threshold decay schedule. The schedule endpoint does not mean that the model must be trained for 400 steps before evaluation. We will clearly distinguish the evaluated checkpoint step from the schedule endpoint in Table 10 and the accompanying explanation to resolve this ambiguity.

References

[1] S. Rui et al. BiWM: Advancing Open-Source Interactive Video World Models with Bidirectional Autoregression. arXiv preprint arXiv:2606.10135, 2026.

[2] M. Liu et al. DriveVA: Video Action Models are Zero-Shot Drivers. ECCV, 2026.

thanks for the rebuttal, my question are mostly resolved.

Comments on the paper

1 comment
Area_Chair_zhPfOpenReview ↗

Discussion

Dear Reviewers, The authors have posted their rebuttal. Please review their responses and engage in the discussion.

Best, AC

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

Related papers

Discussion (0)

Sign in to comment.