acceptodds
Under review as a conference paper at ICLR 2027

WorldRIDE: Efficient Reward-Guided Distillation for Autoregressive World Models

Abstract

Autoregressive world models must combine low-latency generation with visual fidelity and reliable action control. Few-step distillation reduces sampling cost, but treating reward alignment as a subsequent post-training stage introduces another optimization loop. We present WorldRIDE, an efficient reward-guided distillation framework that improves a four-step autoregressive student during distillation itself, without additional post-distillation training. WorldRIDE combines three designs. Ordered reward-guided distillation gates action preferences by group-level visual quality and permits geometric refinement only among action-safe, action-matched candidates. Multi-anchor rollout reuses a single autoregressive trajectory to construct local candidate groups at multiple temporal anchors, amortizing prefix generation across training targets. Locally self-normalized distillation converts within-group rewards into detached, unit-mean weights on the distribution-matching update, retaining unweighted distillation when the quality gate fails. On WBench, WorldRIDE achieves an Overall score of 77.1. On VBench, our four-step model outperforms a 60-step causal model on all three aggregate scores, improving Total, I2V, and Quality by 0.526, 0.767, and 0.286 points, respectively. These results show that reward-guided distillation can improve world-model quality while reducing denoising steps, without a separate stage to recover performance after distillation.

Primary area:
generative models
Keywords:
Interactive World Model, Disttilation

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_MKY7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose WorldRIDE, a framework for distilling world models into efficient interactive streaming world models without additional post-distillation training. The method consists of three main designs: (1) Ordered reward-guided distillation, which applies quality gating before action rewards and enables geometry rewards only on action-valid comparisons. (2) Multi-anchor rollout, which reuses prefix generation to sample local candidates at multiple temporal anchors. (3) Locally self-normalized distillation, which normalizes reward weights within each local candidate group to have unit mean. Experiments on WBench and VBench demonstrate improvements in overall benchmark performance over SANA-WM Streaming.

Critical strengths
  • The paper is well structured and easy to follow.
  • The qualitative results are clearly presented and allow for easy comparison.
  • The proposed solutions are straightforward.
Critical weaknesses
  • Missing comparisons with closely related baselines. One main concern is the lack of direct comparison with methods like Reward Forcing (Re-DMD). Re-DMD already incorporates rewards into distribution matching distillation through reward-weighted updates, which is similar to the core idea of WorldRIDE. An adapted Re-DMD baseline using the same SANA-WM backbone and settings would help identify the improvements more clearly.
  • Insufficient ablations for ordered reward guidance. The results in Table 3 are rather mixed. In particular, removing both action and geometry rewards only reduces the WBench Overall score from 77.1 to 76.9, although the improvement in Perspective Consistency is more noticeable. More importantly, the current ablations do not clarify whether the proposed ordering of rewards is essential. A comparison against a naive combination of quality, action, and geometry rewards without hierarchical gating would help validate the first contribution. In addition, the paper lacks controlled ablations isolating multi-anchor rollout and locally self-normalized weighting, making it difficult to attribute the observed improvements to the individual proposed components.
  • Unclear computational efficiency compared to baselines. Multi-anchor rollout indeed saves computation compared to generating independent full trajectories. However, this does not necessarily imply improved training efficiency over standard distillation. WorldRIDE processes multiple candidates per student update, requiring additional sampling, reward-model inference, and DMD score evaluations. A compute-budget-matched comparison against SANA-WM Streaming and other relevant baselines would be more convincing. The authors should report total training GPU-hours, including rollout generation, reward evaluation, and distillation updates, to better support the efficiency claims.
Questions for authors

See the "Critical Weaknesses section".

Minor weaknesses:

  • Source code is not provided in the submission, which may hinder reproducibility.
3 comments
AuthorsOpenReview ↗

Response to Reviewer MKY7 (Part 1/3): Same-Backbone Comparisons with Reward Forcing / Re-DMD (W1)

Thank you sincerely for your thoughtful, constructive review and for recognizing the clear structure of our paper, and the clarity of our qualitative visual comparisons.

First, we sincerely apologize to you for the delay in posting our response. Two practical factors required substantial time during the rebuttal window: (1) compute constraints, as we trained multiple full-resolution (1280×704) controlled baselines—including two same-backbone Re-DMD baselines and orthogonal ablations—from scratch and evaluated 160 one-minute closed-loop trajectories per model; and (2) the upstream deprecation of WBench's earlier VLM judge (doubao-lite-0215), which required re-evaluating all original baselines alongside our new controlled variants under upstream WBench's current official qwen3.8-max protocol for a strictly fair comparison. Under this unified protocol, all 16 deterministic GPU metrics are bit-identical, and WorldRIDE's Overall gain over SANA-WM Streaming remains invariant at +1.4 (75.7 → 77.1 previously vs. 72.5 → 73.9 under qwen3.8-max).

We organize our response into three parts: Part 1/3 (W1: Same-Backbone Comparisons with Reward Forcing / Re-DMD), Part 2/3 (W2: Controlled Ablations of Ordered Gating, Multi-Anchor Rollout & Local Self-Normalization), and Part 3/3 (W3 & Minor: Total Training GPU-Hours, Compute-Budget Comparison & Code Open-Sourcing).


1. Direct Same-Backbone Comparisons with Reward Forcing (Re-DMD) (W1)

To directly evaluate WorldRIDE against Reward Forcing (Re-DMD) under identical conditions, we implemented and trained two controlled Re-DMD baselines on the exact same SANA-WM Streaming backbone, training dataset (112,573 windows at 1280×704), Causal ODE initialization, and 4-step schedule (Table 1):

Table 1: Same-backbone controlled comparisons against two-stage RL and both faithful & 3-reward adapted Re-DMD baselines (unified WBench & 60s SANA-WM 80-Scene).

4-Step Baseline (Same SANA-WM Backbone & Data)Paradigm (Anchors / Norm / Reward)WBench Overall ↑SegCont ↑NavTraj ↑60s VBench Total (%) ↑60s Camera RotErr (°) ↓
SANA-WM Streaming (official 4-step)K=1 / Unweighted / None72.591.881.969.0722.72
Unweighted DMD (same training data & init)K=1 / Equal (w=1) / None72.899.481.368.8020.33
Self-Forcing + WorldCompass RL (2-stage)Chunk RL (G=12, top/bot 3)72.890.582.068.6719.39
Single-Anchor Quality Re-DMD (faithful Re-DMD)K=1 / Global EMA / HPSv373.494.981.367.6918.34
Single-Anchor 3-Reward Re-DMD (adapted Re-DMD)K=1 / Global EMA / Q+A+Geo73.692.481.368.6319.33
WorldRIDE Full (Ours)K=3 / Group Softmax / Ordered73.998.182.669.1416.58
  • Two fair same-backbone Re-DMD implementations: (1) Single-Anchor Quality Re-DMD faithfully implements original Re-DMD (Lu et al.) using single-anchor rollout (K=1K=1), cross-batch Global EMA weighting, and HPSv3 visual quality reward; (2) Single-Anchor 3-Reward Re-DMD directly tests your suggested adaptation by equipping Re-DMD's K=1K=1 + Global EMA framework with all three of WorldRIDE's quality, action, and geometry rewards.

  • Comparison with faithful Quality Re-DMD (73.4 vs. 73.9): Optimizing single-anchor visual quality alone lacks action and 3D structural supervision—lagging behind WorldRIDE Full in NavTraj (81.3 vs. 82.6), Perspective Consistency (61.7 vs. 67.0, -5.3), and SegCont (94.9 vs. 98.1), while suffering a -1.45% collapse in 60s closed-loop VBench Total (67.69% vs. 69.14%) and higher camera rotation error (18.34° vs. 16.58°).

  • Comparison with adapted 3-Reward Re-DMD (73.6 vs. 73.9): Even when given the exact same three rewards, Re-DMD's single-anchor supervision (K=1K=1), Global EMA normalization, and ungated combination cause severe chunk-boundary discontinuity (SegCont 92.4, -5.7 below WorldRIDE Full 98.1), lower Causal Fidelity (45.7 vs. 49.6, -3.9), and degrade 60s VBench Total below un-distilled Streaming (68.63% < 69.07%) with 19.33° rotation error (25.71° on the hard split). WorldRIDE Full outperforms both Re-DMD variants and two-stage WorldCompass RL across every metric above.

AuthorsOpenReview ↗

Response to Reviewer MKY7 (Part 2/3): Controlled Ablations of Ordered Gating, Multi-Anchor & Local Normalization (W2)

2. Controlled Ablations of Ordered Gating, Multi-Anchor Rollout & Local Self-Normalization (W2)

To attribute gains cleanly to each individual design and clarify the progression in Table 3, Table 2A isolates each of our three contributions on the identical SANA-WM backbone and training dataset (paired with a 192-group / 384-clip live counterfactual gate audit), and Table 2B tracks the step-by-step effect of adding Action and Geometry rewards on long-horizon and 3D metrics.

Table 2A: Single-factor controlled ablations of Contributions 1–3 & 192-group (384-clip) live counterfactual gate audit.

Isolated FactorControlled 4-Step Variant (Same Backbone & Data)Anchors / Norm / GateWrong-Action Trap False Upweight ↓WBench Overall ↑SegCont ↑DynDeg ↑Setting ↑60s VBench Total (%) ↑
Baseline ref.SANA-WM Streaming (official 4-step)K=1 / — / ——72.591.898.771.769.07
K=1 equal-weightUnweighted DMD (same training data)K=1 / Equal (w=1) / ——72.899.497.571.968.80
Contrib. 2 (Multi-anchor)w/o Action + Geometry (multi-anchor equal)K=3 / Equal (w=1) / ——73.698.799.471.769.23
Contrib. 1 (Ungated sum)Multi-Anchor + Ungated Flat SumK=3 / Group Softmax / Flat Sum25.0%74.0*97.594.9 (-1.9)70.468.58
Contrib. 3 (Global norm)Multi-Anchor + Global EMAK=3 / Global EMA / Ordered3.1%74.1*98.782.9 (-13.9!)70.967.42
Full Method (Ours)WorldRIDE FullK=3 / Group Softmax / Ordered3.1%73.998.196.872.369.14

*Denotes short-clip static consistency inflation from under-motion (Global EMA suffers a -13.9 collapse in Dynamic Degree, dropping NavTraj to 81.1 [lowest overall] and crashing 60s VBench Total to 67.42%).

Table 2B: Clarifying Table 3—step-by-step gains of Action and Geometry rewards (fixed K=3, M=2 multi-anchor setup).

Stepwise Variant (Fixed K=3, M=2)Reward HierarchyWBench Overall ↑Perspective ↑60s Camera RotErr (°) ↓60s Revisit PSNR (dB) ↑
w/o Action + Geometry (equal-weight multi-anchor)Equal (w=1)73.662.715.8115.08
w/o Geometry (+ quality-gated Action)Q → Action73.562.613.69 (-13.4%)14.94
WorldRIDE Full (+ conditional Geometry, Ours)Q → A → Geo73.967.0 (+4.4)16.5815.26 (+0.32 dB)
  • Isolating Contribution 2 (Multi-Anchor Shared-Prefix) & why "w/o Action + Geometry" already scores high: Because HPSv3 serves strictly as a binary group feasibility gate in Eq. (2), disabling Action and Geometry yields uniform unit weights (w=1w=1). Thus, comparing Unweighted DMD (K=1,w=1K=1, w=1) to w/o Action + Geometry (K=3,w=1K=3, w=1) strictly isolates multi-anchor rollout: it lifts WBench from 72.8 to 73.6 (+0.8), boosts 60s Revisit PSNR from 14.31 to 15.08 dB, and eliminates single-anchor chunk breaks (SegCont 92.4/94.9 → 98.1 in Table 1), while sharing the prefix spine speeds up 6-branch sampling 6.2× (6,552 → 1,063 ms) and saves 2.86 GB VRAM vs. independent rollouts.

  • Isolating Contribution 1 (Ordered Gating vs. Naive Ungated Sum): Holding K=3,M=2K=3, M=2 and Group Softmax fixed, combining Quality + Action + Geometry without gating falsely upweights sharp wrong-action/static clips 25.0% of the time on 192 live groups; Ordered Gating cuts this 8× to 3.1% (0.0% leakage on degraded-quality motion traps). In end-to-end training, Ungated Flat Sum dampens motion (DynDeg -1.9), drops Setting Adherence to 70.4 (-1.9, lowest overall), and degrades 60s VBench Total below un-distilled Streaming (68.58% < 69.07%).

  • Isolating Contribution 3 (Local Group Softmax vs. Global EMA): Holding K=3,M=2K=3, M=2 and Ordered Gating fixed, Global EMA clips/saturates 51.0%–72.7% of weights (0.0% for Group Softmax) because low-motion scenes systematically earn higher raw scores—inducing severe motion-freeze hacking (DynDeg: 96.8 → 82.9, NavTraj: 81.1, 60s VBench: 67.42%).

  • Step-by-step Action & Geometry gains (Table 2B): Once multi-anchor rollout saturates short 5s WBench paths, 60-second (960-frame) closed-loop trajectories (160 videos, 800 revisit pairs) expose clear orthogonal gains: adding Action cuts 60s RotErr by -13.4% to 13.69° (-39.7% vs. Streaming 22.72°), and adding conditional Geometry boosts WBench Perspective Consistency by +4.4 (62.6 → 67.0) and lifts 60s Revisit PSNR to 15.26 dB (surpassing the 60-step teacher's 14.98 dB).

AuthorsOpenReview ↗

Response to Reviewer MKY7 (Part 3/3): Total Training GPU-Hours, Compute-Budget Comparison & Code Release (W3, Minor)

3. Total Training GPU-Hours, Compute-Budget Comparison & Code Open-Sourcing (W3 & Minor)

To support our efficiency claims with complete end-to-end accounting, Table 3A reports CUDA-synchronized per-step latency (including rollout, candidate decoding, reward inference, and distillation updates), peak trainer VRAM, and 1280×704 streaming inference FPS on NVIDIA H200; Table 3B reports formal end-to-end 1280×704 wall-clock time and total active H200 GPU-hours on one 1×8 NVIDIA H200 node.

Table 3A: Per-step latency (including rollout, decode, reward & DMD updates), peak VRAM, and 1280×704 streaming inference speed (NVIDIA H200, CUDA-synchronized).

Method & Training Paradigm (Same SANA-WM Backbone)Step Time (s/step) ↓Peak Trainer VRAM (GB) ↓720p Infer FPS ↑ (DiT / Stage-1)
SANA-WM Chunk-Causal (60-step Teacher)——1.56 / 1.41
Self-Forcing (unrewarded standard distillation)5.8439.1735.06 / 10.99
Self-Forcing + WorldCompass RL5.84 + 20.8039.17 / 35.2434.90 / 10.97
Global Full-Seq Re-DMD (M=2)20.0654.0735.01 / 10.91
Independent Rollouts (6 separate prefixes)16.3856.9335.01 / 10.91
WorldRIDE (Shared-Prefix + Detached, Ours)15.23 (-26.8% vs. RL)54.0735.01 / 10.91

Table 3B: Formal 1280×704 end-to-end training wall-clock & total active GPU-hours on a 1×8 NVIDIA H200 node (including rollout generation, CausalVAE decoding, HPSv3 + DA3 evaluation & distillation updates).

Method (1280×704, 1×8 H200 Node)Stages (Post-ODE)Wall-Clock Time ↓Active H200 GPU-h ↓Speedup vs. SF + WorldCompass RL ↑
Self-Forcing (unrewarded distillation)1 Stage12.7 h76.2(Unrewarded baseline)
Self-Forcing + WorldCompass RL2 Stages41.6 h (12.7h + 28.9h)307.5 (332.9 node)1.00× (2-stage RL baseline)
WorldRIDE (Ours)1 Stage29.2 h233.61.42× faster (-30% time, -24% GPU-h)
  • Clarifying efficiency relative to standard unrewarded distillation & step breakdown (Table 3A): We fully agree with your observation that processing multiple candidates is not cheaper than unrewarded standard distillation (Self-Forcing / SANA-WM Streaming: 5.84 s/step, 12.7 h / 76.2 H200-h), and we will revise Section 1 and Section 3.2 to state explicitly that our training efficiency claim is relative to reward-aligned world-model pipelines rather than unrewarded DMD. Amortized over the 5:1 critic-to-generator schedule, WorldRIDE's 15.23 s/step includes shared-prefix + 6-branch rollout, local CausalVAE decoding (3.31 s/step amortized), frozen HPSv3 + DA3 scoring (0.94 s/step amortized; depth reprojection <0.2%), and detached DMD updates—making it 26.8% faster per step than WorldCompass RL (20.80 s/step) and 24.1% faster than Global Full-Seq Re-DMD (20.06 s/step), while saving 6.13 GB VRAM over a 4-step differentiable unroll.

  • Superior quality at a lower total GPU-hour budget than two-stage Streaming + WorldCompass RL (Table 3B & Table 1): Counting all rollout generation, candidate decoding, reward inference, and policy/DMD updates (sharing the one-time Causal ODE initialization of 1.3 h / 10.4 H200-h), scaling Self-Forcing (Streaming) with Stage-2 WorldCompass RL consumes 41.6 h (307.5 active H200 GPU-h) yet plateaus at 72.8 WBench (SegCont 90.5, 60s VBench 68.67%). Single-stage WorldRIDE completes in 29.2 h (233.6 active H200 GPU-h)—achieving higher performance across WBench (73.9), SegCont (98.1), 60s VBench (69.14%), and 60s RotErr (16.58°) using 30% less wall-clock time and 24% fewer GPU-hours (1.42× faster), with zero inference overhead (35.01 FPS DiT).

  • Source code & reproducibility (Minor): Due to our internal code-release approval process, we are temporarily unable to upload our code repository at this stage; however, we unequivocally commit to open-sourcing the complete codebase (training, multi-anchor rollout, HPSv3 + DA3 reward service, counterfactual gate audit, and WBench / 80-Scene evaluation scripts, configurations, and checkpoints) upon publication.

Thank you again for your valuable feedback, which has substantially improved our paper. Please let us know if you have any further questions during the discussion period.

Reviewer_Eu7HOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes WorldRIDE, a reward-guided distillation framework for streaming autoregressive world models that integrates capability optimization directly into distribution matching distillation. The key motivation is that visual quality, action following, and 3D consistency have conditional dependencies that should be considered when constructing reward supervision. To address this, the authors introduce ordered reward guidance, which uses a visual-quality gate to enable action-based preferences and further activates geometry rewards only when candidate videos exhibit sufficiently accurate and comparable action execution. The framework also introduces multi-anchor rollout, which generates local candidate groups from shared autoregressive prefixes to reduce redundant sampling, and locally self-normalized distillation, which converts relative rewards into normalized weights for DMD updates. Experiments on WBench and VBench-I2V show that the four-step WorldRIDE student improves over the four-step SANA-WM Streaming baseline.

Critical strengths
  1. Well-motivated reward design for interactive world models. WorldRIDE proposes a reasonable pipeline, i.e., considering visual quality, action following, and 3D consistency.

  2. Efficient integration of reward guidance into the distillation process. WorldRIDE incorporates reward feedback directly into DMD through locally self-normalized weighting, avoiding a separate RL post-training stage.

  3. Promising empirical improvements with few-step generation. WorldRIDE demonstrates improvements over the four-step SANA-WM Streaming baseline on both WBench and VBench-I2V.

Critical weaknesses
  1. Insufficient ablation studies to validate the ordered reward design. The paper lacks comparisons with independently combined rewards or reward guidance without gating. Therefore, it is unclear whether the improvements come from the proposed ordered reward structure or simply from adding reward signals to DMD.

  2. Limited empirical evidence for improvements in action following and geometric consistency. Although the method targets these capabilities, Table 3 shows that gains are mainly concentrated in Perspective Consistency. The full model slightly underperforms the ablated variants in Navigation and Geometric Consistency, raising questions about the effectiveness of the proposed reward guidance.

  3. Insufficient analysis of the efficiency gains from multi-anchor rollout. Although multi-anchor rollout reduces redundant sampling, the paper lacks direct comparisons with independent rollouts in terms of training time, memory usage, and computational overhead, making its practical efficiency gains unclear.

Questions for authors

Please primarily refer to the section Critical Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Eu7H (Part 1/3): Ordered Reward Design vs. Ungated Combinations & Pure DMD (W1)

We sincerely thank you for the positive evaluation and for highlighting our well-motivated reward hierarchy, our efficient integration of reward feedback directly into DMD.

First, we sincerely apologize for the delay in posting our response. Two practical factors required substantial time during the rebuttal window: (1) compute constraints, as we trained multiple full-resolution (1280×704) controlled baselines and ungated/unweighted ablations from scratch and evaluated 160 one-minute closed-loop trajectories per model; and (2) the upstream deprecation of WBench's earlier VLM judge (doubao-lite-0215), which required re-evaluating all original baselines alongside our new controlled variants under upstream WBench's current official qwen3.8-max protocol for a strictly fair comparison. Under this unified protocol, all 16 deterministic GPU metrics are bit-identical, and WorldRIDE's Overall gain over SANA-WM Streaming remains invariant at +1.4 (75.7 → 77.1 previously vs. 72.5 → 73.9 under qwen3.8-max).

We organize our response into three parts: Part 1/3 (W1: Ordered Reward Design vs. Ungated Combinations & Pure DMD), Part 2/3 (W2: Empirical Evidence for Action Following & 3D Geometric Consistency), and Part 3/3 (W3: Efficiency Gains of Multi-Anchor Shared-Prefix vs. Independent Rollouts).


1. Validating Ordered Reward Design vs. Ungated Combinations & Pure DMD (W1)

To separate the effect of ordered conditional gating from (a) pure DMD without rewards and (b) independently combined rewards without gating, Table 1 compares controlled 4-step variants on the identical SANA-WM Streaming backbone and f19 data (112,573 windows at 1280×704), paired with a 192-group (384-clip) live counterfactual gate audit (HPSv3 + DA3).

Table 1: Ordered gating vs. pure DMD and ungated reward combinations (unified WBench, 60s SANA-WM 80-Scene & 192-group audit).

4-Step Student Variant (Same SANA-WM Backbone)Reward OrganizationWrong-Action Trap False Upweight ↓WBench Overall ↑SegCont ↑DynDeg ↑Setting ↑60s VBench Total (%) ↑
SANA-WM Streaming (official 4-step)Unweighted (K=1)—72.591.898.771.769.07
Unweighted DMD (same f19 data & init)No Reward (w=1, K=1)—72.899.497.571.968.80
Single-Anchor 3-Reward Re-DMDUngated Flat Sum (K=1, EMA)25.0%73.692.496.272.468.63
Multi-Anchor + Ungated Flat SumUngated Flat Sum (K=3, Grp)25.0%74.0*97.594.9 (-1.9)70.468.58
WorldRIDE Full (Ours)Ordered Gate (Q → A → Geo, K=3)3.1% (8× lower)73.998.196.872.369.14

*Denotes short-clip static consistency inflation from dampened motion (DynDeg -1.9), which degrades Setting Adherence (70.4, lowest overall) and 60s closed-loop VBench Total (68.58%).

  • Gains do not come from DMD or training data alone (Unweighted DMD vs. WorldRIDE Full): On the exact same f19 data and Causal ODE initialization, Unweighted DMD reaches only 72.8 on WBench (-1.1 below WorldRIDE Full 73.9), 14.31 dB on 60s Revisit PSNR (-0.95 dB below 15.26 dB), and drops 60s VBench Total to 68.80%, confirming that our gains require structured reward guidance rather than DMD alone.

  • Direct mechanism audit (192 shared-prefix groups, 384 live-decoded clips): When HPSv3 + Action + Geometry are combined via an ungated flat sum, sharp static or wrong-action candidates are falsely upweighted 25.0% of the time because low-motion clips naturally earn higher raw quality and depth-reprojection scores. WorldRIDE's Ordered Gate cuts this 8× to 3.1% and achieves 0.0% false-positive leakage on degraded-quality motion traps (vs. 6.2% without quality gating).

  • End-to-end training necessity (Ungated Flat Sum vs. Ordered Gate): Simply adding all three rewards to DMD without gating degrades world-model behavior in both single- and multi-anchor regimes: Single-Anchor 3-Reward Re-DMD suffers cross-chunk breaks (SegCont 92.4), while holding K=3K=3 multi-anchor rollout and Group Softmax completely fixed and replacing Ordered Gating with Ungated Flat Sum dampens motion (DynDeg -1.9), drops Setting Adherence to 70.4 (-1.9, lowest overall), and degrades 60s VBench Total below the un-distilled baseline (68.58% < 69.07%). Only Ordered Gating (WorldRIDE Full) preserves dynamic motion and setting adherence while improving 60s closed-loop quality (69.14%).

AuthorsOpenReview ↗

Response to Reviewer Eu7H (Part 2/3): Empirical Evidence for Action Following & 3D Geometric Consistency (W2)

2. Empirical Evidence for Action Following & 3D Geometric Consistency (W2)

To clarify why short-clip WBench Navigation and Geometric Consistency appear close across reward ablations while both rewards drive clear, orthogonal improvements on their targeted capabilities, Table 2 evaluates the exact step-by-step progression across both unified WBench and 60-second (960-frame) closed-loop camera trajectories (SANA-WM 80-Scene, 160 videos, 800 revisit frame pairs).

Table 2: Step-by-step action following and 3D geometric consistency gains across 4-step student variants (unified WBench & 60s SANA-WM 80-Scene).

4-Step Student Variant (Same SANA-WM Backbone)Reward ConfigurationWBench Overall ↑NavTraj (5s) ↑Perspective ↑60s Camera RotErr (°) ↓60s Revisit PSNR (dB) ↑60s Revisit LPIPS ↓
SANA-WM Streaming (official 4-step)Unweighted (K=1)72.581.949.622.7214.260.5697
Unweighted DMD (same f19 data & init)Equal (w=1, K=1)72.881.360.420.3314.310.5444
w/o Action + Geometry (equal-weight multi-anchor)Equal (w=1, K=3)73.682.962.715.8115.080.5372
w/o Geometry (+ quality-gated Action)Q → Action (K=3)73.582.762.613.6914.940.5402
WorldRIDE Full (+ conditional Geometry)Q → A → Geo (K=3)73.982.667.016.5815.260.5343
  • Why 5-second WBench NavTraj and Geometric Consistency saturate: WBench evaluation clips are short (~5s single-pass paths); once multi-anchor shared-prefix distillation (w/o Action + Geometry) suppresses cross-chunk drift, NavTraj rises from 81.3 (Unweighted DMD) to 82.9—already approaching the 60-step teacher ceiling (84.5, so 82.6–82.9 reflects ±0.3 ceiling variation). Meanwhile, WBench's Geometric Consistency measures adjacent-frame local epipolar smoothness (which favors smaller viewpoint changes), whereas our bidirectional depth-reprojection reward optimizes cross-boundary and chunk-span 3D structure under camera motion—directly captured by WBench Perspective Consistency and long-horizon closed-loop revisit fidelity.

  • Direct Action Following gain (w/o Action + Geometry → w/o Geometry): On 60-second (960-frame) closed-loop trajectories (SANA-WM 80-Scene, 160 videos) where pose drift compounds across 24 autoregressive chunks, adding the quality-gated Action reward cuts 60s camera rotation error RotErr by -13.4% (15.81° → 13.69°)—a -39.7% error reduction over official SANA-WM Streaming (22.72°) and -32.7% over Unweighted DMD (20.33°).

  • Direct 3D Geometric Consistency gain (w/o Geometry → WorldRIDE Full): Activating the conditional Geometry reward among action-matched candidates boosts WBench Perspective Consistency by +4.4 (62.6 → 67.0) and lifts WBench Overall to 73.9. Across 800 revisit frame pairs over 60-second closed-loop trajectories, it raises Revisit PSNR from 14.94 to 15.26 dB (surpassing even the 60-step teacher's 14.98 dB) and lowers perceptual revisit error Revisit LPIPS to 0.5343 (with mild regularization on aggressive turn amplitude, keeping RotErr 16.58° -27.0% below Streaming).

AuthorsOpenReview ↗

Response to Reviewer Eu7H (Part 3/3): Multi-Anchor Shared-Prefix vs. Independent Rollouts Efficiency (W3)

3. Efficiency Gains of Multi-Anchor Shared-Prefix vs. Independent Rollouts (W3)

To directly quantify the practical efficiency gains of Multi-Anchor Shared-Prefix Rollout against Independent Rollouts and related training paradigms in terms of computational overhead, memory usage, training time, and inference speed, Table 3A reports CUDA-synchronized per-step profiling and 1280×704 streaming inference FPS on NVIDIA H200; Table 3B reports formal end-to-end 1280×704 distributed training wall-clock time and active H200 GPU-hours on one 1×8 NVIDIA H200 node.

Table 3A: Direct H200 CUDA-synchronized comparison of Shared-Prefix vs. Independent Rollouts (sampling identical K×M = 6 local target chunks) and related paradigms.

Rollout & Training Paradigm (Same SANA-WM Backbone)Spine (ms)6-Branch Sampling (ms) ↓Step Time (s/step) ↓Peak Trainer VRAM (GB) ↓720p Infer FPS ↑ (DiT / Stage-1)
SANA-WM Chunk-Causal (60-step Teacher)————1.56 / 1.41
Self-Forcing (single-anchor, unrewarded)2,97205.8439.1735.06 / 10.99
Self-Forcing + WorldCompass RL—G=12 rollouts5.84 + 20.8039.17 / 35.2434.90 / 10.97
Global Full-Seq Re-DMD (M=2 full-seq decode)5,9440 (2× full-seq)20.0654.0735.01 / 10.91
Independent Rollouts (6 separate prefixes)3,0196,55216.3856.9335.01 / 10.91
WorldRIDE w/ 4-Step Diff. Unroll (no detach)3,0151,06318.0860.2035.01 / 10.91
WorldRIDE (Shared-Prefix + Detached, Ours)3,0151,063 (6.2× faster)15.23 (-26.8% vs. RL)54.07 (-2.86 / -6.13 GB)35.01 / 10.91

Table 3B: Formal 1280×704 end-to-end training wall-clock & active GPU-hours on a 1×8 NVIDIA H200 node.

Method (1280×704, 1×8 H200 Node)Stages (Post-ODE)Wall-Clock Time ↓Active H200 GPU-h ↓Speedup vs. SF + WorldCompass RL ↑
Self-Forcing (unweighted DMD)1 Stage12.7 h76.2(Unrewarded baseline)
Self-Forcing + WorldCompass RL2 Stages41.6 h (12.7h + 28.9h)307.5 (332.9 node)1.00× (2-stage RL baseline)
WorldRIDE (Ours)1 Stage29.2 h233.61.42× faster (-30% time, -24% GPU-h)
  • Computational overhead (6.2× faster branch sampling & 50% less decode/reward cost): To sample the exact same K×M=3×2=6K\times M = 3\times 2 = 6 local target chunks across depths, Independent Rollouts must generate 6 separate prefixes from scratch (6,552 ms branch sampling), whereas reusing a single shared autoregressive spine (3,015 ms) cuts 6-branch sampling to 1,063 ms (6.2× faster). Moreover, decoding and scoring only local anchor windows (3.31 s/step amortized CausalVAE decode + 0.94 s/step amortized HPSv3 + DA3, with depth reprojection <0.2%) halves decode/reward overhead vs. Global Full-Seq Re-DMD, reducing total step latency by 24.1% (15.23 vs. 20.06 s/step) and 26.8% vs. WorldCompass RL (20.80 s/step).

  • Memory usage (-2.86 GB vs. Independent Rollouts; -6.13 GB vs. 4-step unroll): Reusing the shared prefix KV-cache avoids holding multiple independent rollout states simultaneously, saving 2.86 GB peak trainer VRAM over Independent Rollouts (54.07 vs. 56.93 GB); combining this with detached exit-step recomputation saves 6.13 GB peak VRAM (54.07 vs. 60.20 GB) and 2.85 s/step over a 4-step differentiable unroll.

  • End-to-end training wall-clock & zero inference overhead: Sharing the same one-time Causal ODE initialization (1.3 h / 10.4 H200-h), single-stage WorldRIDE completes in 29.2 h (233.6 active H200-h) on one 1×8 H200 node—1.42× faster (-30% wall-clock time and -24% active GPU-hours) than two-stage Self-Forcing + WorldCompass RL (41.6 h / 307.5 H200-h). At inference, all teachers and reward models are discarded, achieving 35.01 FPS DiT / 10.91 FPS Stage-1 (17.22 GB VRAM)—zero overhead vs. Self-Forcing and 22.5× faster DiT than the 60-step teacher.

We thank you again for the supportive review and constructive comments, and we remain happy to address any further questions during the discussion period.

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

WorldRIDE incorporates reward feedback into the distillation of a camera-controlled world model into a four-step autoregressive student, avoiding a separate reward post-training stage. The method combines three components: conditional reward gating based on visual quality, action following, and geometric evidence; multi-anchor sampling that constructs local candidate groups along a shared autoregressive trajectory; and detached, locally normalized reward weights that modulate distribution-matching distillation updates. Compared with the four-step Streaming baseline, WorldRIDE improves WBench aggregate performance and several consistency metrics. Long-horizon results show improvements in revisit fidelity and camera rotation, alongside regressions in translation error and temporal quality retention.

Critical strengths
  1. The reward hierarchy has a well-motivated, task-specific rationale. Restricting geometric refinement to visually acceptable candidates with adequate action execution and valid geometric evidence addresses a meaningful failure mode: static outputs may receive high geometric consistency scores despite ignoring motion commands. This conditional organization considers dependencies between objectives rather than simply summing independent rewards, making it a consequential design choice for interactive world modeling.
  2. Consistency improvements are supported by quantitative evaluations without the refiner. Without refinement, WorldRIDE improves WBench Overall from 73.6 to 74.6. These results demonstrate that some benefits are present in the student outputs themselves, rather than arising exclusively from post-processing or selected qualitative examples.
Critical weaknesses
  1. The contribution of the proposed reward hierarchy is not clearly established. In Table 3, removing both Action and Geometry rewards already achieves 76.9 Overall, compared with 77.1 for the full method and 75.7 for Streaming. Moreover, the full method does not improve Navigation or Geometric Consistency over the reward ablations; its clearest benefit is in Perspective Consistency. These results leave the source of most of the improvement over Streaming unresolved. The paper also lacks controlled comparisons isolating ordered gating, multi-anchor rollout, and local normalization, making it difficult to assess the necessity of the three claimed contributions.
  2. The claimed training efficiency is not demonstrated in terms of actual computational cost. Avoiding a separate RL stage does not necessarily reduce total training cost. WorldRIDE introduces candidate decoding, HPSv3 and DA3 evaluation, and target recomputation, but reports neither actual training time nor GPU-hours. Clip-sampling counts and training-step counts alone cannot establish a computational advantage over standard distillation followed by RL. This is a material gap because efficiency is central to the paper’s motivation and contribution.
Questions for authors
  1. Clarification of gain attribution: Does the “w/o Action + Geometry” variant retain all other WorldRIDE settings? What differences from Streaming account for its improvement from 75.7 to 76.9, and how should we interpret the lack of Navigation and Geometric Consistency gains from the full reward configuration?
  2. Evidence for the proposed mechanisms: Can you provide controlled comparisons of ordered versus ungated reward weighting, multi-anchor versus independent rollouts, and local normalization versus standard reward weighting? These comparisons would directly affect the assessment of the claimed contributions.
3 comments
AuthorsOpenReview ↗

Response to Reviewer NLMe (Part 1/3): Gain Attribution & Role of Action/Geometry Rewards (W1, Q1)

We sincerely thank you for the insightful reviews. First, we sincerely apologize for the delay in posting our response. Two practical factors required substantial time during the rebuttal window: (1) compute constraints, as we trained multiple full-resolution (1280×704) controlled baselines and orthogonal ablations from scratch and evaluated 160 one-minute closed-loop trajectories per model; and (2) the upstream deprecation of WBench's earlier VLM judge (doubao-lite-0215), which required re-evaluating all original baselines alongside our new controlled variants under upstream WBench's current official qwen3.8-max protocol for a strictly fair comparison. Under this unified protocol, all 16 deterministic GPU metrics are bit-identical, and WorldRIDE's Overall gain over SANA-WM Streaming remains invariant at +1.4 (75.7 → 77.1 previously vs. 72.5 → 73.9 under qwen3.8-max).

We organize our response into three parts: Part 1/3 (W1 & Q1: Gain Attribution & Role of Action/Geometry Rewards), Part 2/3 (W1 & Q2: Controlled Mechanism Isolations), and Part 3/3 (W2: Training Time, GPU-Hours, Peak Memory & Inference Speed).


1. Gain Attribution of "w/o Action + Geometry" & Role of Action / Geometry Rewards (W1, Q1)

Table 1: Step-by-step gain attribution across 4-step student variants (unified WBench & 60s SANA-WM 80-Scene).

4-Step Student Variant (Same SANA-WM Backbone)Anchors / Weighting SignalWBench ↑SegCont ↑NavTraj (5s) ↑Perspective ↑60s PSNR (dB) ↑60s RotErr (°) ↓
SANA-WM Streaming (official 4-step)K=1 / Unweighted72.591.881.949.614.2622.72
Unweighted DMD (same data & init)K=1 / Equal (w=1)72.899.481.360.414.3120.33
w/o Action + Geometry (equal-weight multi-anchor)K=3, M=2 / Equal (w=1)73.698.782.962.715.0815.81
w/o Geometry (+ quality-gated Action)K=3, M=2 / Q → Action73.597.582.762.614.9413.69
WorldRIDE Full (+ conditional Geometry)K=3, M=2 / Q → A → Geo73.998.182.667.015.2616.58
  • Q1(a) Exact specification of "w/o Action + Geometry": Yes—it retains all other WorldRIDE settings (identical data, Causal ODE init, 4-step schedule, and K=3K=3 multi-anchor shared-prefix rollout with M=2M=2 local branches/anchor). Because HPSv3 serves strictly as a binary group feasibility gate gQgrp(k)∈{0,1}g_Q^{\mathrm{grp}}(k)\in\{0,1\} in Eq. (2) rather than a standalone continuous reward, disabling Action and Geometry sets Ak,jord≡0A^{\mathrm{ord}}_{k,j}\equiv 0 and yields uniform unit weights w~k,j≡1\widetilde{w}_{k,j}\equiv 1, making it an equal-weight multi-anchor shared-prefix DMD control.

  • Q1(b) Source of the improvement from Streaming to "w/o Action + Geometry": Single-anchor Unweighted DMD (K=1K=1) on the exact same data and ODE init yields only +0.3 over official Streaming (72.5 → 72.8; 60s PSNR flat at 14.26 → 14.31 dB), ruling out data or schedule differences. The primary jump (72.8 → 73.6 WBench Overall, 81.3 → 82.9 NavTraj, and 14.31 → 15.08 dB 60s Revisit PSNR) comes directly from multi-anchor shared-prefix rollout (K=3,M=2K=3, M=2, Contribution 2), which supervises local targets across multiple autoregressive depths to suppress cumulative exposure bias.

  • Q1(c) Why 5s WBench NavTraj/Geom saturate while Action & Geometry drive long-horizon & 3D perspective gains:

    1. WBench clips are short (~5s single-pass paths); once multi-anchor distillation stabilizes cross-chunk drift, NavTraj (82.9) already approaches the 60-step teacher ceiling (84.5). Meanwhile, WBench's Geometric Consistency measures adjacent-frame local epipolar smoothness (favoring smaller viewpoint shifts), whereas our bidirectional depth-reprojection reward optimizes cross-boundary and chunk-span 3D structure under camera motion—captured directly by Perspective Consistency on WBench and closed-loop revisit / pose error over 60-second (960-frame) trajectories.
    2. On 60-second closed-loop trajectories (SANA-WM 80-Scene, 160 videos, 800 revisit pairs): adding Action (w/o Geometry) cuts 60s camera rotation error RotErr by -13.4% to 13.69° (-39.7% vs. Streaming 22.72°); adding conditional Geometry (WorldRIDE Full) boosts WBench Perspective Consistency by +4.4 (62.6 → 67.0) and lifts 60s Revisit PSNR to 15.26 dB (surpassing even the 60-step teacher's 14.98 dB), with mild regularization on aggressive turns (RotErr 16.58°, still -27.0% below Streaming).
AuthorsOpenReview ↗

Response to Reviewer NLMe (Part 2/3): Controlled Mechanism Isolations (W1, Q2)

2. Controlled Mechanism Isolations

To directly assess the necessity of each of our three claimed contributions on the identical SANA-WM Streaming backbone and dataset, Table 2A isolates Ordered Gating vs. Ungated Sum (Q2a), Multi-Anchor vs. Single-Anchor (Q2b), and Local Group Softmax vs. Standard Global EMA Weighting (Q2c), paired with a 192-group (384-clip) live counterfactual gate audit; Table 2B compares Shared-Prefix vs. Independent Rollouts (Q2b).

Table 2A: Controlled mechanism ablations & 192-group (384-clip) live counterfactual gate audit.

Isolated MechanismControlled VariantAnchors / Norm / GateWrong-Action Trap False Upweight ↓WBench Overall ↑SegCont ↑DynDeg ↑Setting ↑60s VBench Total (%) ↑
Baseline ref.SANA-WM Streaming (official)K=1 / — / ——72.591.898.771.769.07
Q2(b) Single-anchorSingle-Anchor Quality Re-DMDK=1 / Global EMA / Qual64.1%73.494.998.171.067.69
Q2(b) Single-anchorSingle-Anchor 3-Reward Re-DMDK=1 / Global EMA / Flat25.0%73.692.496.272.468.63
Q2(c) Global normMulti-Anchor + Global EMAK=3 / Global EMA / Ordered3.1%74.1*98.782.9 (-13.9!)70.967.42
Q2(a) Ungated sumMulti-Anchor + Ungated Flat SumK=3 / Group Softmax / Flat Sum25.0%74.0*97.594.9 (-1.9)70.468.58
Full MethodWorldRIDE Full (Ours)K=3 / Group Softmax / Ordered3.1%73.998.196.872.369.14

*Asterisks denote under-motion static inflation on short 5s clips (Global EMA suffers a -13.9 collapse in Dynamic Degree, dropping NavTraj to 81.1 [lowest overall] and crashing 60s VBench to 67.42%).

Table 2B: Q2(b) Shared-Prefix vs. Independent Full Rollouts for identical K×M = 6 target chunks (1× H200, CUDA-synchronized).

Rollout Strategy (Identical 6 Target Chunks)Spine (ms)6-Branch Sampling (ms) ↓Amortized Step (s/step) ↓Peak Trainer VRAM (GB) ↓
Independent Rollouts (6 separate prefixes)3,0196,55216.3856.93
Multi-Anchor Shared-Prefix (Ours)3,0151,063 (6.2× faster)15.2354.07 (-2.86 GB)
  • Q2(a) Ordered Gating vs. Ungated Reward Weighting: On 192 live-decoded shared-prefix anchor groups (384 clips scored by frozen HPSv3 + DA3), Ungated Flat Sum (HPSv3 + Action + Geometry without gating) falsely upweights sharp wrong-action/static clips 25.0% of the time; WorldRIDE's Ordered Gate cuts this 8× to 3.1% (and achieves 0.0% leakage on degraded-quality motion traps vs. 6.2% without quality gating). In full training, Ungated Flat Sum slightly dampens motion (DynDeg -1.9), inflating short-clip static consistency (74.0*) at the expense of the lowest Setting Adherence (70.4, -1.9) and a 60s VBench Total drop below the un-distilled baseline (68.58% < 69.07%).

  • Q2(b) Multi-Anchor Shared-Prefix vs. Single-Anchor & Independent Rollouts: Against single-anchor Re-DMD (K=1K=1, Table 2A), multi-anchor supervision (K=3K=3) eliminates chunk-boundary breaks (SegCont: 92.4 / 94.9 → 98.1). Against generating 6 independent prefixes from scratch for the exact same 6 targets (Table 2B), reusing the shared spine accelerates branch generation by 6.2× (6,552 → 1,063 ms) and saves 2.86 GB peak VRAM.

  • Q2(c) Local Self-Normalization vs. Standard Global EMA Weighting: Under standard cross-batch Global EMA normalization, 51.0%–72.7% of candidate weights clip/saturate (0.0% for Group Softmax) because low-motion scenes systematically earn higher raw quality/geometry scores—triggering severe motion-freeze hacking (DynDeg: 96.8 → 82.9, NavTraj: 81.1, 60s VBench: 67.42%). Group Softmax normalizes strictly among siblings sharing the exact same scene, prefix, and action command, removing cross-scene static bias by construction.

AuthorsOpenReview ↗

Response to Reviewer NLMe (Part 3/3): Training Time, GPU-Hours, Memory & Inference Speed (W2)

3. Actual Training Time, GPU-Hours, Peak Memory & Inference Speed (W2)

To establish computational cost directly in terms of wall-clock latency, GPU-hours, memory footprint, and inference throughput, Table 3A reports CUDA-synchronized per-step training latency, peak trainer VRAM, and steady-state 1280×704 streaming inference FPS on NVIDIA H200; Table 3B reports formal end-to-end 1280×704 distributed training wall-clock time and active H200 GPU-hours on one 1×8 NVIDIA H200 node.

Table 3A: Controlled per-step latency, peak trainer VRAM, and 1280×704 streaming inference speed (NVIDIA H200, CUDA-synchronized).

Method (Same SANA-WM Backbone)Step Time (s/step) ↓Peak Trainer VRAM (GB) ↓720p Infer FPS ↑ (DiT / Stage-1)
SANA-WM Chunk-Causal (60-step Teacher)——1.56 / 1.41
Self-Forcing5.8439.1735.06 / 10.99
Self-Forcing + WorldCompass RL5.84 + 20.8039.17 / 35.2434.90 / 10.97
Global Full-Seq Re-DMD (M=2)20.0654.0735.01 / 10.91
WorldRIDE (Ours)15.23 (-26.8% vs. RL)54.0735.01 / 10.91

Table 3B: Formal 1280×704 end-to-end training wall-clock & active GPU-hours on a 1×8 NVIDIA H200 node.

Method (1280×704, 1×8 H200 Node)Stages (Post-ODE)Wall-Clock Time ↓Active H200 GPU-h ↓Speedup vs. SF + WorldCompass RL ↑
Self-Forcing (unweighted DMD)1 Stage12.7 h76.2(Unrewarded baseline)
Self-Forcing + WorldCompass RL2 Stages41.6 h (12.7h + 28.9h)307.5 (332.9 node)1.00× (2-stage RL baseline)
WorldRIDE (Ours)1 Stage29.2 h233.61.42× faster (-30% time, -24% GPU-h)
  • Per-step overhead & memory efficiency (Table 3A): Amortizing local CausalVAE candidate decoding (3.31 s/step amortized) and frozen HPSv3 + DA3 evaluation (0.94 s/step amortized; bidirectional depth reprojection is <0.2%) over the 5:1 critic-to-generator schedule makes WorldRIDE 26.8% faster per step than WorldCompass RL (15.23 vs. 20.80 s/step, where G=12 candidate rollouts and 13 CausalVAE decodes dominate RL wall-clock) and 24.1% faster than full-sequence Global Re-DMD (20.06 s/step). Meanwhile, detached exit-step target recomputation saves 6.13 GB peak VRAM (54.07 vs. 60.20 GB) and 2.85 s/step compared with a 4-step differentiable unroll.

  • Total training cost & zero inference overhead (Table 3B): Sharing the same one-time Causal ODE initialization (1.3 h / 10.4 H200-h), single-stage WorldRIDE completes in 29.2 h (233.6 active H200-h) on a 1×8 H200 node—1.42× faster (-30% wall-clock time and -24% active GPU-hours) than two-stage Self-Forcing + WorldCompass RL (41.6 h / 307.5 active H200-h), while avoiding a separate post-distillation RL pipeline and improving both short-horizon consistency and 60s revisit fidelity. At inference, all teachers and reward models are discarded, achieving 35.01 FPS DiT / 10.91 FPS Stage-1 (17.22 GB VRAM)—zero overhead vs. Self-Forcing and 22.5× faster DiT than the 60-step teacher.

We thank you again for the constructive feedback, which has substantially strengthened the empirical rigor of our paper, and we welcome any further questions during the discussion period.

Comments on the paper

1 comment

Dear Area Chairs and Reviewers,

Alongside our individual responses, we provide an anonymous page with synchronized qualitative video comparisons and consolidated benchmark tables (note: although SANA-WM is a compact 2.6B backbone with bounded base capacity, the visual gains brought by WorldRIDE remain clear):

https://worldride-rebuttal-41835.vercel.app/rebuttal/index.html

Thank you again for your time and valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.