acceptodds
Withdrawn from ICLR 2027

Mass Forcing: On-Policy Mass-Aware Distillation for Autoregressive Video Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

32 views

Abstract

Distribution Matching Distillation (DMD) has emerged as a promising paradigm for real-time video synthesis, distilling pretrained diffusion models into few-step generators. However, its reverse-KL objective is mode-seeking and can collapse the student's conditional video distribution onto a narrow subset of modes, leading to reduced sample diversity, diminished temporal dynamics, and color oversaturation. Although existing methods primarily mitigate this issue via coverage-oriented student initialization, we observe that such initial coverage is transient: across three different initializations, the fraction of dynamic videos drops sharply during subsequent DMD training, indicating that the unchanged mode-seeking objective, rather than the initialization, drives this degradation. Motivated by this training-time degradation, we propose Mass Forcing, a probability-mass-aware framework that incorporates coverage-oriented supervision throughout on-policy DMD training via complementary local and set-level objectives. First, at the local level, Mass Forcing estimates relative teacher-to-student density ratios across rollouts sharing a student-generated prefix through latent-space score integration and uses these estimates to rebalance conditional DMD updates, amplifying the training influence of rollouts underrepresented relative to the teacher. Second, recognizing that mismatched mode probabilities do not necessarily induce strong local score corrections, Mass Forcing aligns generated rollout sets with a reweighted reference distribution defined by the same mass estimates, providing cross-rollout supervision beyond individual gradient reweighting. Experiments on MovieGen and VBench demonstrate that Mass Forcing improves video fidelity and sustains temporal dynamics throughout training, achieving the highest Total scores among compared methods and a Dynamic Degree of 83% after 1,000 generator updates, versus 12-32% for the baselines, and introducing no additional inference-time computational overhead.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Mode Seeking, Distributional Matching Distillation, Autoregressive Video Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UuFeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Mass Forcing, a probability-mass-aware on-policy distillation framework. Given multiple student continuations sampled from a shared generated prefix, it estimates their relative teacher-to-student density ratios by integrating teacher–student score differences along latent-space paths. Experiments on MovieGen and VBench show improvements in aggregate quality and Dynamic Degree over several autoregressive and distilled video-generation baselines. The method does not modify the four-step inference pipeline or introduce additional inference-time computation.

Critical strengths
  • The paper identifies an important limitation of current autoregressive DMD pipelines. The observation that improved initialization does not prevent temporal dynamics from degrading during continued DMD training is practically important. The training trajectories across Self Forcing, Causal Forcing, and Causal-rCM provide useful evidence that initialization alone may not resolve the problem.

  • The proposed method contains a meaningful objective-level contribution. Unlike methods that only modify student initialization, Mass Forcing changes the training signal throughout on-policy optimization. Estimating relative density ratios from existing teacher and fake-score networks, and combining local mass weighting with set-level matching, is technically interesting. The method is also practical at inference because it preserves the original generator, sampler, and latency.

Critical weaknesses
  • The paper does not establish that the observed decline in Dynamic Degree corresponds to mode-coverage collapse. Dynamic Degree measures whether a video contains noticeable motion, but it does not measure conditional sample diversity, mode recall, or agreement with the teacher’s mode probabilities. A model can achieve high Dynamic Degree while generating only one type of motion, excessive camera movement, or unstable visual changes. Therefore, the central explanation that reverse-KL causes conditional mode collapse remains plausible but insufficiently validated.

  • A critical baseline is missing, preventing clear attribution of the gains to mass-aware supervision. Mass Forcing changes not only the loss but also the training data path: it samples (K) independent continuations from a shared prefix and applies conditional-chunk DMD, whereas Self Forcing uses a holistic video path. The paper does not report the corresponding uniform-weight baseline with (K=4), (hkh_k=1), and (γ\gamma=0). Consequently, it is unclear whether the improvements come from density-ratio weighting and MMD or simply from grouped multi-rollout conditional training.

  • The reliability and limitations of the density-ratio estimator are insufficiently validated. The estimator integrates approximate score differences along straight paths in a high-dimensional latent space. These paths may cross low-density regions where both teacher and fake-score networks are unreliable. Moreover, because all (K=4) candidates are sampled from the student, the method cannot reliably identify a teacher-supported mode that is absent or extremely unlikely under the current student. The current ablations show that weight assignment matters, but they do not demonstrate that the estimated weights accurately represent teacher-to-student probability ratios.

  • The experimental evidence is incomplete for the claimed streaming-video setting. The main evaluation focuses on short videos and automatic VBench metrics. The paper lacks long-horizon rollout evaluation, human assessment of motion realism, uncertainty estimates across training runs, and training-cost analysis. These omissions are important because higher Dynamic Degree may reflect excessive or low-quality motion rather than improved temporal diversity.

Questions for authors
  • Density-ratio validation: Can the authors directly validate the estimated relative density ratios, for example, through a controlled Gaussian-mixture experiment or by comparing their rankings with an independently trained density-ratio discriminator?

  • Missing baseline: What are the results of an unweighted grouped-rollout baseline using the same sampled chunk, shared prefix, (K=4) continuations, training budget, and replay procedure, but with (hkh_k=1) and (γ\gamma=0)?

  • Mode-coverage evaluation: Can the authors report direct conditional-diversity or mode-recall metrics across multiple seeds for the same prompt, rather than using Dynamic Degree as the main proxy for mode coverage?

  • Teacher–student discrepancy: In Table 1, Mass Forcing achieves substantially higher Dynamic Degree than Wan2.1. Is the reported Wan2.1 model the same 14B teacher used during training? If so, how should a student that exceeds the teacher’s dynamic frequency be interpreted as better matching the teacher distribution?

  • Estimator stability: How sensitive are the estimated ratios to the integration path, number of quadrature points, noise timestep, medoid anchor, and fake-score estimation error? The endpoint-only ablation does not fully establish that the line integral provides a reliable density-ratio estimate.

3 comments
AuthorsOpenReview ↗

Answers for Reviewer UuFe (1/2)

Thank you for the detailed and insightful review, and for recognizing a "meaningful objective-level contribution." All new results below will be added to the revision in blue.

W1/Q3. DD vs. mode coverage. We now measure conditional diversity and mode recall directly, across seeds for the same prompt. For each of 32 prompts, we generate 32 videos with different seeds and compare them, in VideoMAE feature space, with 32 videos from the 14B teacher for the same prompt. Diversity is the mean pairwise L2 distance within a prompt. Precision and recall use kk-NN with k=3k=3 and the teacher videos as reference.

ModelDiversity ↑Precision ↑Recall ↑
Teacher (14B), new seeds (upper bound)0.8650.7740.830
ODE Init0.6870.0100.000
Self Forcing0.5560.3510.000
Mass Forcing0.7870.4550.045

From the same initialization, Self Forcing loses diversity (0.687 → 0.556), which supports the reverse-KL explanation, while Mass Forcing gains it (0.787). A single type of motion would lower diversity, and excessive camera movement or unstable changes would lower teacher-relative precision. Instead, both rise over Self Forcing (diversity 0.556 → 0.787, precision 0.351 → 0.455). Recall is near zero for every 1.3B student, including the initialization, and Mass Forcing is the only one above zero. In the blind user study (W4), Mass Forcing is also most often ranked first on motion naturalness and imaging quality. We treat DD only as a dynamics metric.

W2/Q2. Matched baseline. We trained the baseline you describe. It uses the same sampled chunk, shared prefix, K=4K=4 continuations, training budget (1,000 updates), and replay as Mass Forcing, and also the same initialization, critic, optimizer, and seed. Only hk=1h_k=1 and γ=0\gamma=0 differ.

MethodMass weightingMMDMovieGen TotalMovieGen DDVBench TotalVBench DDDD at update 100 → 500 → 1,000
Self Forcing✗✗81.7563.4183.4961.1166 → 36 → 13
Uniform conditional-chunk✗✗81.5562.4981.9860.3265 → 33 → 11
w/o MMD✓✗82.4065.0083.8177.7865 → 69 → 65
Mass Forcing✓✓83.3883.3584.9888.8965 → 77 → 83

Total and DD use the official checkpoint for Self Forcing, the best checkpoint (update 400) for the uniform baseline, and update 800 for w/o MMD. The last column is from our own training runs.

The uniform baseline collapses like Self Forcing (DD 65% → 11%) and does not improve Total even at its best checkpoint. Grouped multi-rollout conditional training alone therefore explains neither gain. Mass weighting halts the decline and adds +0.85/+1.83 to Total (MovieGen/VBench). MMD then lifts DD to 83% and adds another +0.98/+1.17.

W3/Q1. Estimator reliability. We ran the controlled Gaussian-mixture experiment you suggested, with 4 components in 2 and 32 dimensions and known density ratios. Teacher weights are uniform and student weights are [0.70, 0.20, 0.09, 0.01]. As in training, ratios are estimated at σ\sigma ∈ [0.20, 0.65] with K=4K=4 groups. We report how often two samples of a group are ordered correctly (%, chance 50), with exact scores (a perfect fake score) and with a fake score learned as in Mass Forcing.

σ\sigmaExact scores, d=2Exact scores, d=32Learned fake score, d=2Learned fake score, d=32
0.2096.196.277.384.3
0.42597.297.382.689.5
0.6598.598.693.990.6
  • Low-density paths: with exact scores, straight paths crossing the low-density gaps between modes still order 96-99% of pairs correctly. With the learned fake score, 77-94% are still ordered correctly, and with q = p the weights remain near uniform (L1 ≤ 0.012). Any remaining error has a bounded effect per update: with λmax⁡=0.25\lambda_{\max}=0.25, each hk∈[0.75,1.75]h_k\in[0.75,1.75] (Appendix B.4).
  • Absent modes: we agree. All candidates come from the student, so such modes cannot be up-weighted. Our claim is that coverage is preserved during training, and we will add this limitation to the paper.
  • Discriminator: we chose the mixtures because they give exact ground truth. A discriminator adds its own estimation error.
AuthorsOpenReview ↗

Answers for Reviewer UuFe (2/2)

Q4. Teacher vs. Table 1. No. The Wan2.1 row in Table 1 is the 1.3B model, not the 14B teacher. We will label it Wan2.1-1.3B in the table. We also do not read a higher DD as closer matching to the teacher. DD is a thresholded motion rate, not a distance between distributions, and the training target includes CFG (scale 3.0). The relevant test is the teacher-relative comparison in W1/Q3, where Mass Forcing is the closest student to the teacher on all three metrics.

Q5. Stability. On our model, we change one factor at a time for 96 groups of K=4K=4 samples in active cells (8 prompts × 2 chunks × 2 prefixes × 3 noise levels) and compare with the default estimator:

ChangeSpearman rank correlation ↑L1 change in weights hh ↓
3 → 9 Simpson points0.9540.016
Rotated anchor0.9290.030
Resampled query noise0.9350.019

The median rank correlation is 1 in all three cases, and in at least 90% of groups the ranking of the four samples changes by at most one adjacent swap. A two-segment path through another sample matches the rotated-anchor values to within 5×10−55\times 10^{-5} nats.

The ratio at each σ\sigma is between σ\sigma-smoothed densities, so its value depends on σ\sigma by design. What matters is that the ordering is correct at each σ\sigma, which the mixtures confirm at all three levels (W3/Q1). For fake-score error, the mixtures show that a learned fake score still orders 77-94% of pairs correctly (chance 50%) and keeps the weights near uniform when q = p. Its effect on each update is also bounded (W3/Q1).

W4. Long videos. We add 60-second results on VBench-Long with the standard VBench prompts. Reward Forcing is itself a long-video method and uses its native inference. The others use Infinity-RoPE.

VBench-Long (60 s)Self ForcingCausal ForcingReward ForcingMass Forcing
Quality79.3183.9579.9784.03
Semantic69.0571.2074.7078.95
Total77.2681.4078.9283.01

Mass Forcing is best on all three. In the last 5 s window (20 prompts × 2 seeds), its DD is 78% vs. 9% for Self Forcing, while its imaging quality only drops from 69.9 to 67.8.

W4. Human assessment. We ran a blind user study. 116 participants each ranked 10 sets of four videos (8 short, 2 long, from 100 prompts) on motion naturalness, dynamic degree, and imaging quality, with video order randomized. Share ranked first (%, chance 25):

Ranked first (%)Self ForcingCausal ForcingReward ForcingMass Forcing
Short videos
Motion naturalness1.214.05.079.8
Dynamic degree1.229.53.465.8
Imaging quality2.47.810.279.6
Long videos
Motion naturalness1.724.19.564.7
Dynamic degree1.345.34.748.7
Imaging quality2.210.322.465.1

Mass Forcing leads on naturalness and quality as well as dynamics (Friedman p < 0.001), so its extra motion is not judged excessive. Side-by-side 5 s and 60 s comparisons are also in the supplementary demo video.

W4. Training runs. We trained Mass Forcing again with a different training seed (300 GPU-hours per run). Total is 83.54/84.80 (MovieGen/VBench) vs. 83.38/84.98 for the reported run, and DD is 82.36/88.77 vs. 83.35/88.89. The two runs differ by at most 0.18 in Total and 1.0 in DD, and both exceed every baseline in Table 1 in Total and DD on both benchmarks. We will add this run to the revision. The effect is also consistent across the other separately trained runs. All four runs without mass weighting (Self Forcing, Causal Forcing, and Causal-rCM in Figure 1, and the uniform baseline) end at 11-32% DD, while the runs with it end at 65-83% (W2/Q2).

W4. Cost. Mass Forcing costs 2.6× the FLOPs of Self Forcing per generator update. Ratio estimation and MMD add only 1% over the uniform baseline, because the Simpson midpoint needs one extra 1.3B residual call rather than a 14B teacher query. Inference is unchanged.

Presentation. We will add a four-step overview of the estimator to Section 2.2 and keep the gating details in Appendix C.

AuthorsOpenReview ↗

Paper Revision Uploaded

Thank you again for your review. We have uploaded a revised version of the paper. All changes are highlighted in blue.

  • W1/Q3 (DD vs. mode coverage). Appendix G reports conditional diversity, precision, and recall against the 14B teacher in Table 13. Section 3.1 now states that we treat DD as a measure of temporal dynamics and not of mode coverage.
  • W2/Q2 (matched baseline). Section 3.4 now opens with this baseline, and Appendix D.5 gives the full comparison in Table 12.
  • W3/Q1 (estimator reliability). Appendix H.1 reports the Gaussian-mixture study in Table 14. It also reports the magnitude error, which is why the claim is restricted to the ordering of the weights.
  • Q4 (teacher vs. Table 1). The row is now labeled Wan2.1-1.3B.
  • Q5 (stability). Appendix H.2 reports the one-factor-at-a-time study in Table 15.
  • W4 (long videos). Appendix I, Table 16 and Figure 9.
  • W4 (human assessment). Appendix J, Table 17.
  • W4 (training runs). Appendix K.1 compares the two training seeds in Table 18.
  • W4 (cost). Appendix K.2.
  • Presentation. Section 2.2 now includes a four-step overview of the estimator, and the gating details remain in Appendix C.2.
  • Limitations. Section 5 now states that modes the student has already lost cannot be recovered.

The attachment contains a demo video with side-by-side comparisons at 5 and 60 seconds.

If any concerns remain, we would be glad to address them before the discussion period ends.

Reviewer_LbqSOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper tries to address a long-standing problem in the field of diffusion distillation where DMD causes the dynamics to drop significantly and color saturation. It utilizes two techniques to mitigate the problem which include relative teacher-to-student density ratios and cross-rollout supervision. Experiments show that the proposed method achieves SOTA results across most dimensions of various benchmarks.

Critical strengths
  1. the motivation is well justified which tries to solve a critical problem in DMD distillation.
  2. the approach is grounded in distribution misalignment which includes teacher-student density analysis and cross-rollout alignment.
  3. the proposed method achieves SOTA results across various benchmarks.
Critical weaknesses
  1. The results are only validated for 5 second short generation which makes the results less convincing.
  2. There are no visual results attached in the appendix or anywhere, it's hard to judge the quality of generations from images.
Questions for authors

Besides the questions in weakness, how does the proposed method compare to other methods such as Reward-Forcing [1] or Infinity-RoPE [2].

[1] Lu, Yunhong, et al. "Reward forcing: Efficient streaming video generation with rewarded distribution matching distillation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026. [2] Yesiltepe, Hidir, et al. "Infinity-rope: Action-controllable infinite video generation emerges from autoregressive self-rollout." Proceedings

2 comments
AuthorsOpenReview ↗

Answers for Reviewer LbqS

Thank you for the review. The new results below will be added to the revision in blue.

W1. 5-second evaluation. We respectfully note that 5 s is the standard protocol of every method we compare with. The paper studies the loss of dynamics during DMD training, and none of its claims depend on video length. We nonetheless add 60-second results on VBench-Long with the standard VBench prompts. Reward Forcing is itself a long-video method and uses its native inference. The others use Infinity-RoPE.

VBench-Long (60 s)Self ForcingCausal ForcingReward ForcingMass Forcing
Quality79.3183.9579.9784.03
Semantic69.0571.2074.7078.95
Total77.2681.4078.9283.01

Mass Forcing is best on all three. In the last 5 s window (20 prompts × 2 seeds), it keeps DD at 78% vs. 9% for Self Forcing, and its imaging quality drops only from 69.9 to 67.8.

W2. Visual results. The submission is not without visual results. It contains frame sequences in Figures 1, 4, 6, and 7 and in Appendix F, and the motion can be read from the changes between frames. We now also provide a demo video in the supplementary material, with side-by-side comparisons with Self Forcing and Causal Forcing at 5 s, and with Reward Forcing added at 60 s.

We also ran a blind user study. 116 participants each ranked 10 sets of four videos (8 short, 2 long, from 100 prompts) on motion naturalness, dynamic degree, and imaging quality, with video order randomized. Share ranked first (%, chance 25):

Ranked first (%)Self ForcingCausal ForcingReward ForcingMass Forcing
Short videos
Motion naturalness1.214.05.079.8
Dynamic degree1.229.53.465.8
Imaging quality2.47.810.279.6
Long videos
Motion naturalness1.724.19.564.7
Dynamic degree1.345.34.748.7
Imaging quality2.210.322.465.1

Mass Forcing is most often ranked first on all three criteria, for both short and long videos (Friedman p < 0.001).

Q. Reward Forcing and Infinity-RoPE. Both will be cited in the revision. Reward Forcing is compared at 5 s and 60 s, in the demo video, and in the user study. Infinity-RoPE is applied to Self Forcing, Causal Forcing, and ours at 60 s.

  • Reward Forcing combines EMA-Sink, which targets long-horizon streaming and is orthogonal to our work, with Re-DMD, which weights DMD by a VLM motion reward. Mass Forcing instead weights rollouts by teacher-to-student density ratios and needs no external reward. At 60 s, the setting it is designed for, Reward Forcing scores 78.92 VBench-Long Total vs. 83.01 for ours (W1). In the user study, Mass Forcing is ranked first far more often on every criterion (49-80% vs. 3-22% of sets, W2). Under our 5 s protocol, the official checkpoint scores 84.20 Total and 72.22 DD on VBench, vs. 84.98 Total and 88.89 DD for ours.
  • Infinity-RoPE is a training-free inference technique for Self-Forcing-style models, not a distillation method, so it is complementary to our objective rather than a competing baseline. With Infinity-RoPE applied to both, Mass Forcing scores 83.01 VBench-Long Total at 60 s vs. 77.26 for Self Forcing, and keeps last-window DD at 78% vs. 9% (W1). Since the inference technique is the same, this gap comes from training.
AuthorsOpenReview ↗

Paper Revision Uploaded

Thank you again for your review. We have uploaded a revised version of the paper. All changes are highlighted in blue.

  • W1 (5-second evaluation). Appendix I adds 60-second results on VBench-Long in Table 16, where Mass Forcing is highest in Quality, Semantic, and Total.
  • W2 (visual results). Figure 9 shows 60-second rollouts of all four methods at matched timestamps, from the same prompt and the same seed. Appendix J reports the blind user study with 116 participants in Table 17. The attachment contains a demo video with side-by-side comparisons at 5 and 60 seconds.
  • Q (Reward Forcing and Infinity-RoPE). Both are cited in Section 4. Reward Forcing is now a baseline in Table 1 at 5 seconds and in Table 16 at 60 seconds.

Both weaknesses you raised concern the scope of the evaluation rather than the method, and both are now addressed with new experiments. We would appreciate it if you could let us know whether this resolves your concerns, or what remains open.

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

Mass Forcing addresses temporal-dynamics degradation during continued DMD training for few-step autoregressive video generation. It estimates relative teacher-to-student density ratios across shared-prefix continuations and uses them for conditional DMD reweighting and rollout-set MMD matching. Experiments on MovieGen and VBench show improved aggregate quality and sustained dynamics without additional inference-time computation.

Critical strengths
  1. The paper identifies a practically relevant training failure mode. The trajectories across different initializations, together with Appendix D.4, show that temporal dynamics can deteriorate even as visual quality improves.
  2. The complementary local and set-level objectives are well motivated. Removing MMD or shuffling weight assignments degrades performance, providing encouraging evidence for the proposed components. These improvements require no additional inference-time computation.
Critical weaknesses
  1. A matched unweighted baseline is missing. Mass Forcing changes holistic video DMD into conditional-chunk updates with detached prefixes. The ablations do not include an unweighted conditional-chunk baseline under otherwise identical settings. Consequently, the contributions of the proposed mass-aware objectives and the modified training procedure remain insufficiently separated.
  2. The evidence for improved mode coverage remains indirect. Dynamic Degree measures the fraction of generated videos classified as dynamic, but does not directly quantify conditional diversity or mode coverage. Although Appendix B.4 acknowledges the estimator’s approximation limitations, direct validation of the estimated density ratios is missing. The construction and validation criteria for the calibration certificate and prevalidated activation cells in Appendix C.2 are also insufficiently specified.
Questions for authors
  1. Can you report an unweighted conditional-chunk DMD baseline with \lambda_{max}=0 and \gamma=0, under otherwise identical initialization, rollout sampling, critic configuration, and optimization settings? How does it compare with Mass Forcing in final quality and training-time dynamics?
  2. Can you provide direct validation of the estimated mass weights and improvements in conditional mode coverage? How are the calibration certificate and prevalidated activation cells constructed and validated?
2 comments
AuthorsOpenReview ↗

Answers for Reviewer Ajxb

Thank you for the constructive review, and for recognizing "a practically relevant training failure mode." All new results below will be added to the revision in blue.

W1/Q1. Matched unweighted baseline. We trained the baseline you describe. It is identical to Mass Forcing (initialization, chunk sampling, shared prefix, K=4K=4, replay, critic, optimizer, seed, 1,000 updates) except λmax⁡=0\lambda_{\max}=0 (hk=1h_k=1) and γ=0\gamma=0.

MethodMass weightingMMDMovieGen TotalMovieGen DDVBench TotalVBench DDDD at update 100 → 500 → 1,000
Self Forcing✗✗81.7563.4183.4961.1166 → 36 → 13
Uniform conditional-chunk✗✗81.5562.4981.9860.3265 → 33 → 11
w/o MMD✓✗82.4065.0083.8177.7865 → 69 → 65
Mass Forcing✓✓83.3883.3584.9888.8965 → 77 → 83

Total and DD use the official checkpoint for Self Forcing, the best checkpoint (update 400) for the uniform baseline, and update 800 for w/o MMD. The last column is from our own training runs.

  • Final quality: even at its best checkpoint, the uniform baseline trails Mass Forcing by 1.83/3.00 Total (MovieGen/VBench), and it does not improve over Self Forcing.
  • Training-time dynamics: its DD falls from 65% to 11% by update 1,000, just like Self Forcing.

The conditional-chunk procedure alone therefore explains neither gain. Mass weighting halts the decline and adds +0.85/+1.83 to Total. MMD then lifts DD to 83% and adds another +0.98/+1.17.

W2/Q2. Conditional mode coverage. Beyond DD, we now measure coverage directly. For each of 32 prompts, we generate 32 videos with different seeds and compare them, in VideoMAE feature space, with 32 videos from the 14B teacher for the same prompt. Diversity is the mean pairwise L2 distance within a prompt. Precision and recall use kk-NN with k=3k=3 and the teacher videos as reference.

ModelDiversity ↑Precision ↑Recall ↑
Teacher (14B), new seeds (upper bound)0.8650.7740.830
ODE Init0.6870.0100.000
Self Forcing0.5560.3510.000
Mass Forcing0.7870.4550.045

From the same initialization, Self Forcing loses diversity (0.687 → 0.556), as reverse-KL mode seeking predicts. Mass Forcing gains it (0.787) and also raises precision relative to the teacher (0.351 → 0.455), so the extra variety does not come from off-distribution samples. Recall is near zero for every 1.3B student, including the initialization, and Mass Forcing is the only one above zero. We treat DD only as a dynamics metric.

W2/Q2. Validation of the mass weights. We use 4-component Gaussian mixtures in 2 and 32 dimensions, where the density ratios are known. Teacher weights are uniform, and student weights are [0.70, 0.20, 0.09, 0.01]. As in training, ratios are estimated at σ\sigma ∈ [0.20, 0.65] with K=4K=4 groups. We report how often two samples of a group are ordered correctly (%, chance 50), with exact scores (a perfect fake score) and with a fake score learned as in Mass Forcing.

σ\sigmaExact scores, d=2Exact scores, d=32Learned fake score, d=2Learned fake score, d=32
0.2096.196.277.384.3
0.42597.297.382.689.5
0.6598.598.693.990.6

With exact scores, 96-99% of pairs are ordered correctly, which isolates the integration error. With the learned fake score, 77-94% are still ordered correctly, and with q = p, the weights remain near uniform (L1 ≤ 0.012).

W2/Q2. Calibration certificate and activation cells. Both are fixed before DMD training and are not tuned on evaluation results. We will add the full criteria to Appendix C.2. The certificate checks that the fake score, fitted to the initial student, is accurate on unseen prompts and along the integration paths. Each cell (a noise level and chunk) is active only if the teacher uses the conditioning there and the fake score is accurate. Concretely, a cell passes if the teacher's error drops significantly under true vs. shuffled conditioning (bootstrap over 32 probe prompts) and the fake score's error there is within 4× the median over all cells. The certificate passes, and 37 of the 42 cells are active. Elsewhere, the update is exactly the uniform baseline of Q1.

Although both are set at initialization, training stays stable as the student changes. At updates 500 and 1,000, all 37 cells pass again. The fake score also keeps tracking the current student, as in standard DMD. Weighting is weakened or switched off whenever the weights concentrate, and each weight stays in [0.75, 1.75].

AuthorsOpenReview ↗

Paper Revision Uploaded

Thank you again for your review. We have uploaded a revised version of the paper. All changes are highlighted in blue.

  • W1/Q1 (matched unweighted baseline). Section 3.4 now opens with this baseline, and Appendix D.5 gives the full comparison in Table 12.
  • W2/Q2 (coverage beyond DD). Appendix G reports conditional diversity, precision, and recall against the 14B teacher in Table 13.
  • W2/Q2 (validation of the mass weights). Appendix H.1 reports the Gaussian-mixture study in Table 14.
  • W2/Q2 (certificate and activation cells). Appendix C.2 now states the full criteria, together with the re-check during training.

The attachment contains a demo video with side-by-side comparisons at 5 and 60 seconds.

If any concern remains, we would be glad to address it before the discussion period ends.

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

Related papers

Discussion (0)

Sign in to comment.