acceptodds
Under review as a conference paper at ICLR 2027

Memory-Gated Distillation: Mitigating Self-Generated Memory Drift in Autoregressive Diffusion World Model Distillation

Abstract

Autoregressive diffusion world models enable interactive long-horizon video generation, but their sampling cost is incurred at every generated chunk, making low-latency deployment difficult. Few-step distillation from a bidirectional teacher to an autoregressive student is a promising remedy, but rollout distillation becomes unstable over long horizons. We identify self-generated memory drift as a key cause of this instability: early student errors are written into memory and reused as conditioning context, degrading subsequent distillation. We introduce Memory-Gated Distillation (MGD), which addresses this error-accumulation problem through selective memory supervision. MGD uses a tailored chunk-level reliability verifier to decide whether each student-generated chunk should be written into memory, thereby preserving reliable rollout prefixes while filtering corrupted ones. As a result, MGD retains the benefits of on-policy student rollouts while suppressing supervision from corrupted memory contexts, with no additional inference-time cost. Experiments on short- and long-horizon generation demonstrate improved few-step generation quality and a stronger quality-latency tradeoff than prior distillation baselines. Particularly in the challenging 253-frame rollout setting, 4-step MGD nearly matches the 50-step bidirectional teacher on aesthetic quality (0.564 vs. 0.569) and PSNR (13.81 vs. 13.84) with less than 1/10 of sampling time.

Primary area:
generative models
Keywords:
diffusion, world model, distillation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_RLN2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper is based on the observation that if the history is broken during autoregressive rollout, it will introduce future generations with unreliable supervision from teacher models. Thus based on this finding, it proposes a method to mitigate this problem. The way it works is to create a verifier which can evaluate the quality of the generated trunks and mask out the low quality ones during training. This way, the student model will only get high quality supervisions from the teacher model. Through emprical experiments, the paper shows that it's able to achieve competitive results while maintaining low cost.

Critical strengths
  1. The paper tries to address a critical problem in creating long rollouts for autoregressive generation.
  2. The observation the paper bases on is a widely observed among different model families and methods, thus the motivation is well grounded
  3. The proposed method is simple and effective
Critical weaknesses
  1. the paper tries to address the long rollout problem, however it didn't compare with other popular methods other than just the baseline method and context forcing
  2. The paper shares similar idea with [1], however this is not discussed in detail and compared with.
  3. The code is attached in the supplemental material, however, no visual results are added, it's hard to judge the quality
  4. The result is only applied to HY-World

[1] Yang, Ying, et al. "Stableworld: Towards stable and consistent long interactive video generation." arXiv preprint arXiv:2601.15281 (2026).

Questions for authors
  1. can the authors provide comparison with other long-rollout methods
  2. results with other baseline methods beyond HY-World
  3. can the authors provide more visual results and compare these to the results from other methods?
Reviewer_JJVnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is an extension of self-forcing. It proposes a "self-generated memory drift" concept where autoregressive (AR) student may generate out-of-distribution input during roll-out, which hurts the distillation. This work propose Memory-Gated Distillation (MGD) trains a chunk-level verifier to separate reference chunks from student chunks under shared conditioning and it is designed to determine whether to trust the student-generated context or not. Experiments on HY-World 1.5 shows improvement over the ungated baseline.

Critical strengths
  • Clear writing and a practical problem for on-policy AR distillation. The proposed method is simple, and it leaves the backbone and inference unchanged. Two design choices are simple and practical: keeping the fake score ungated, and using the verifier only for selection rather than for adversarial training.
  • Long-horizon behavior close to the teacher at 4 steps. At 253 frames, 4-step MGD is close to the 50-step BI teacher on several perceptual and geometry metrics (aesthetic 0.564 vs. 0.569, LPIPS 0.462 vs. 0.456, translation error 0.096 vs. 0.096), and it avoids the near-static collapse of the ungated baseline (dynamic degree 0.35 vs. 0.02).
Critical weaknesses
  • My main concern is the technical novelty of the work. Using a discriminator as a density-ratio estimator to reject samples is an established idea (e.g., Discriminator Rejection Sampling, Azadi et al., ICLR 2019, which is not cited). Here it is moved from sampling-time rejection to training-time selection of supervised positions, and the mechanism amounts to a classifier, a threshold, and a prefix mask.
  • Missing Baselines: the only training baseline is labeled “Context Forcing,” but it appears to be the authors’ own τ → −∞ setting (Table 3: “No verifier (Context Forcing)”). Its dynamic degree of 0.02 suggests it may be under-tuned. There is no comparison with Self Forcing, Self-Forcing++, Causal Forcing, or Rolling Forcing, all of which the paper cites.
  • The claim of detecting memory-drift is a bit confusing. According to App. A, the verifier does not take the memory latents as input, and its positives are ground-truth chunks rather than teacher samples. It seems the purpose is mainly to check realism of the generated chucks.
Questions for authors
  • How does MGD compare against the existing baselines including Self Forcing, Self-Forcing++, Causal Forcing, or Rolling Forcing?
  • Does the gate’s behavior or the resulting gain change if the verifier receives the memory latents, or is trained with teacher-generated positives? How often, and at which chunk index, does the gate reject over the course of training? More ablation and analysis is appreciated.
Reviewer_WkgKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses self-generated memory drift in few-step distillation of autoregressive diffusion world models. It proposes Memory-Gated Distillation (MGD), which combines a chunk-level verifier with a causal gate mask for selective distribution matching distillation, while alternating updates to the student, fake-score model, and verifier. Four-step generation experiments using HY-World 1.5 on Sekai show improved long-horizon generation quality and dynamics. The auxiliary modules are used only during training and introduce no inference-time overhead.

Critical strengths
  1. A valuable research problem and a targeted approach. The paper highlights the risk that early generation errors are repeatedly reused as conditioning for subsequent chunks, complementing the perspective of reducing the train-test mismatch through on-policy rollouts alone. Equation (10) organizes selective supervision into a contiguous prefix and retains the first chunk as an anchor, while Section 4.4 keeps the fake-score model tracking the full student distribution. The modules have clearly defined roles, and only the student is retained at inference, making the approach practically relevant.

  2. Meaningful improvements in long-horizon quality and dynamics. Under the same four-step inference budget, Table 1 shows that dynamic degree on 253-frame rollouts increases from 0.020 for Context Forcing to 0.350 for MGD. In Table 2, PSNR increases from 11.65 to 13.81, while LPIPS decreases from 0.494 to 0.462. The joint improvement in generation quality and dynamics provides stronger evidence of reduced long-horizon degradation in the evaluated setting than higher consistency alone. Tables 3 and 4 also provide initial ablation support for learned gating and moderate selection.

Critical weaknesses
  1. Limited direct evidence for the memory reliability interpretation. Figure 1 and Tables 1 and 2 show long-horizon degradation and improvements from MGD, consistent with the paper’s motivation. However, the Introduction and Section 4.1 further attribute these gains to avoiding unreliable distillation supervision under memory drift. The paper does not yet provide direct diagnostics of how history quality affects subsequent supervision, or an analysis relating verifier scores to subsequent degradation. Appendix A states that the implemented verifier does not directly take autoregressive memory as input. This simplification does not imply that the verifier is ineffective, but its connection to the memory reliability interpretation in Section 4.2 needs clarification. The current evidence does not sufficiently distinguish this mechanism from generic selection based on conditional generation quality, limiting the support for the central interpretation.

  2. The contribution of learned prefix selection has not been sufficiently isolated. Table 3 supports the effectiveness of learned gating in the evaluated setting. However, the no-verifier, reference-distance gate, and learned gate variants supervise an average of 4.0, 2.1, and 2.6 chunks, respectively. Table 4 likewise changes both gating strength and the extent of supervision. It therefore remains unclear how much of the improvement comes from the learned reliability signal and how much comes from reducing supervision on later chunks. A simple prefix-truncation baseline with a matched training budget and amount of supervision would more directly assess the contribution attributable specifically to learned selection.

Questions for authors
  1. Evidence for memory reliability. Could the authors provide more direct evidence for the proposed memory reliability interpretation, for example through a controlled comparison of histories or an analysis of whether verifier scores predict degradation in subsequent chunks? In light of the simplification described in Appendix A, where the verifier does not directly take memory as input, please clarify how its scores reflect risks to subsequent generation rather than only the quality of the current chunk.

  2. The distinct contribution of learned selection. With the training budget and average number of supervised chunks matched, does MGD still outperform a simple baseline that limits supervision to a fixed or randomly selected prefix length? This comparison would help distinguish the contribution of reliability assessment from that of reducing supervision on later chunks.

If the authors can convincingly clarify and address these issues, providing evidence for the memory reliability interpretation and the distinct contribution of learned prefix selection, I would be willing to raise my rating from 2 to 3.

Reviewer_SjH4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on the few-step distillation of autoregressive diffusion world models. Specifically, it addresses the issue of self-generated memory drift, where errors in student rollouts accumulate in memory and compromise subsequent distillation supervision. To mitigate this, the authors propose Memory-Gated Distillation (MGD), which employs a chunk-level verifier and a causal gate mask to restrict distribution matching distillation to a valid contiguous rollout prefix. Experiments on HY-World 1.5 with Sekai data evaluate 61 and 253 frame generation. Four-step MGD improves perceptual quality and reduces motion collapse compared with Context Forcing. Ablations further support the effectiveness of the learned verifier and a moderate gating threshold.

Critical strengths
  • This paper highlights that on-policy rollouts do not necessarily guarantee reliable distillation supervision. This intuition is well-motivated: accumulated errors in self-generated memory may move the conditioning context outside the teacher’s reliable operating region. It provides a useful motivation for considering supervision quality alongside exposure to student-generated histories.
  • The experiments demonstrate improvements in long-horizon generation with 4-step inference. In the 253-frame setting, MGD enhances perceptual quality over the native 4-step AR model, alleviates motion collapse compared to Context Forcing, and approaches the 50-step bidirectional teacher in terms of aesthetic quality and PSNR.
Critical weaknesses
  • The experiments do not directly validate the claimed memory-drift mechanism. The current experiments infer the role of the chunk-level reliability verifier from improvements in final generation quality. However, these results do not directly establish whether its scores identify memory drift associated with unreliable subsequent supervision. I would suggest adding a controlled history-perturbation experiment to examine whether perturbing a history chunk lowers its verifier score and using the resulting memory increases the teacher’s denoising error on the next chunk. This would provide more direct support for the proposed mechanism.
  • Limited comparisons with representative distillation methods. Context Forcing is the only explicitly identified distillation method baseline, leaving MGD’s advantages over other approaches unclear. I would suggest adding representative methods such as Self Forcing and Self-Forcing++ using the same backbone and conditioning, with comparable training budgets and four-step inference, to provide a broader assessment of the proposed method.
Questions for authors
  • Section 4.2 describes a verifier conditioned on autoregressive memory, whereas Appendix A states that the implemented verifier does not receive memory latents. Could the authors clarify this difference and explain how the verifier assesses memory reliability without memory inputs?
  • Appendix A uses ground-truth chunks as verifier positives, rather than teacher continuations generated under the student’s history. When the student’s history has already diverged from the ground-truth history, how do these positives help the verifier assess memory reliability, rather than simply distinguish real from generated chunks?
  • Could the authors specify the checkpoints used for student initialization and the HY-World 1.5 AR 4-step baseline? In particular, does this baseline use the official four-step distilled checkpoint, or the undistilled AR model with its sampling steps reduced to four?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.