{"id":"336ac8c6-a815-440c-b225-c762c8265d11","arxiv_id":"2412.11706","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"AsymRnR selectively reduces query and key/value tokens in video DiT attention to cut FLOPs and latency by 10 to 30 percent with minor or no VBench score change.","lead":"This paper introduces AsymRnR, a training-free method that speeds up video diffusion transformers by removing redundant tokens in attention, using different reduction rates for query versus key/value features. A generalist reader might care because it reports 1.1x to 1.3x faster video generation on models like HunyuanVideo and CogVideoX without retraining and with roughly unchanged quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VBench per-dimension scores show non-negligible degradation (e.g., Multiple Objects drops 0.10 for CogVideoX-2B), undermining the 'negligible degradation' claim.","rationale":"The reader's weakest assumption concerns the context-agnostic scheduling in Section 3.4. That assumption is partially supported by the paper's own pipeline: a schedule tuned on one arbitrary prompt is then applied to the entire 900-prompt VBench suite and produces the reported aggregate scores, which is a natural transfer test. The quality-degradation concern, by contrast, is directly evidenced in the paper's own Table 11 and challenges the literal wording of the central claim. The aggregate VBench score hides a 0.102 absolute drop on 'Multiple Objects' for CogVideoX-2B, a dimension that is semantically important and likely perceptible. Without multiple seeds or significance testing, the 'negligible' qualifier is not established, and the 'in some cases, even improving it' language could reflect noise rather than a real effect. This is a load-bearing weakness because it targets the quality-preservation pillar of the central claim, not just the plug-and-play convenience. The method's speedup numbers are credible, and the aggregate quality change is small, so a conditional acceptance remains appropriate; the paper should add statistical rigor and qualify its claims about negligible degradation.","tokens_in":21680,"tokens_out":12872,"duration_ms":119611,"concrete_test":"Run the VBench evaluation for CogVideoX-2B baseline and AsymRnR with 5 independent random seeds on the full 900-prompt suite. For each of the 16 dimensions and the aggregate score, compute paired differences with 95% bootstrap confidence intervals. If the 'Multiple Objects' dimension shows a mean drop of 0.05 or more with a confidence interval excluding 0, the 'negligible degradation' claim is not supported for that dimension and should be qualified; if all confidence intervals include 0, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim asserts 'negligible degradation in output quality' (Abstract). The paper's own per-dimension VBench results (Table 11) contradict this on key semantic dimensions. For CogVideoX-2B, the default AsymRnR configuration drops 'Multiple Objects' from 0.6502 to 0.5480 (an absolute drop of 0.102), 'Color' from 0.8776 to 0.8252 (0.052), and 'Scene' from 0.5378 to 0.5029 (0.035), while the aggregate VBench score falls only 0.0092. The aggregate masks large, likely perceptible degradations. No error bars, multiple seeds, or significance tests are reported anywhere in Section 4 or Appendix E. The paper's own Limitation section (Section 5) admits 'visual discrepancies in the generated outputs' yet the Abstract still uses the word 'negligible'. Since 'negligible degradation' is a central component of the claim, this missing statistical support is load-bearing: a user relying on the headline claim could receive substantially worse multiple-object and color fidelity at the reported speedups.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AsymRnR, a training-free token-reduction method for accelerating video diffusion transformers. It asymmetrically reduces query and key/value token sequences before self-attention and restores them afterward, schedules reductions adaptively across blocks and denoising steps based on precomputed similarity statistics, and introduces a matching cache that reuses bipartite matching results across timesteps. Experiments on CogVideoX-2B/5B, Mochi-1, HunyuanVideo, FastVideo-Hunyuan, and AnimateDiff report latency speedups of about 1.1x-1.3x with aggregate VBench score changes near 0.01 or better on several models, plus a combination with the PAB feature-caching method reaching 1.71x. The paper also presents a KL-divergence estimator (Corollary 3.1) as a theoretical motivation for Euclidean-distance matching.","tokens_in":21907,"tokens_out":6163,"duration_ms":54983,"significance":"If the empirical claims hold, AsymRnR is a practical and broadly applicable plug-in: it is training-free, combines with step distillation and feature caching, is evaluated on a standard benchmark with over 900 prompts, and comes with released code. The consistent speedups across five model families, the clear ablations of scheduling, matching cache, similarity metric, and partition stride, and the explicit compatibility discussion are strengths. However, the central claim of 'negligible degradation' is supported only at the aggregate VBench level; the paper's own per-dimension tables show substantial drops on semantic dimensions such as Multiple Objects and Color, and no uncertainty quantification is provided. The theoretical motivation is suggestive but is not a proof that KL-minimizing token reduction preserves attention outputs. Overall the contribution is valuable and the empirical core is defensible, but the quality claim and the generalization claim need to be substantially qualified and better supported.","major_comments":[{"comment":"The abstract's 'negligible degradation' is contradicted by the paper's own per-dimension VBench results. For CogVideoX-2B, the default AsymRnR configuration drops Multiple Objects from 0.6502 to 0.5480 (an absolute decrease of 0.102), Color from 0.8776 to 0.8252 (0.052), and Scene from 0.5378 to 0.5029 (0.035), while the aggregate VBench falls only from 0.8008 to 0.7917 (0.009). The OURS-FAST row shows a comparable Multiple Objects drop from 0.6502 to 0.5686. No error bars, multiple seeds, or significance tests are reported for VBench or latency in Section 4 or Appendix E, so it is impossible to judge whether these per-dimension degradations are reproducible or perceptually meaningful. Because 'negligible degradation' is a headline claim, the paper should either report uncertainty and a per-dimension analysis, or explicitly qualify the claim to something like 'small aggregate VBench change with some per-dimension degradation.'","section":"§4.2, Table 11; Abstract"},{"comment":"The reduction schedule rests on the assumption that per-feature token similarity patterns are 'context-agnostic and can be considered an intrinsic property of the models.' Figure 4 asserts consistency across diverse contents, but no quantitative transfer evidence is given, and Appendix B shows that the similarity thresholds and reduction rates are tuned per model by visual inspection on a small number of prompts. If the similarity map is prompt-dependent, the precomputed schedule will reduce the wrong tokens for unseen prompts, and the claimed plug-and-play advantage disappears. Please provide evidence that a schedule estimated on one prompt set transfers to another, or at least a sensitivity analysis showing that VBench and latency are robust to the threshold choices.","section":"§3.4, Eq. (5); Appendix B"},{"comment":"The theoretical foundation has several issues. Corollary 3.1 applies to covariance-stationary sequences sampled from two distributions P and P', but the reduction operates on a single deterministic token sequence, and the selected tokens depend on empirical distances computed from the same sequence, so the stationarity/independence condition is not satisfied. In addition, the additive term in Eq. (2), written 'log l/l′ − 1', does not match the estimator in Appendix Eq. (10); for the stated sample sizes the term should be log(l/(l′−1)), and as written it can even be negative. Finally, minimizing D_KL(P'||P) between token distributions is not shown to be equivalent to preserving the attention output softmax(QK^⊤)V; the paper motivates but does not prove this connection. Since the abstract advertises 'theoretical foundations,' this section should be corrected or explicitly labeled as heuristic motivation.","section":"§3.1, Corollary 3.1, Eq. (2); Appendix A"}],"minor_comments":[{"comment":"The word 'metioned' should be 'mentioned'.","section":"§3.5"},{"comment":"The sentence 'the results are summarized in Table 1' should refer to Table 3, which reports the PAB combination.","section":"§4.2, 'Integration with Feature Caching'"},{"comment":"The sentence 'The corresponding qualitative results are provided in Table 4' should say 'quantitative results' or point to the actual figure, since Table 4 contains numbers rather than qualitative comparisons.","section":"§4.2, 'Integration with UNet-based video diffusion models'"},{"comment":"The header contains the typo 'MUTLIPLE OBJECTS'; it should be 'MULTIPLE OBJECTS'.","section":"Table 11"},{"comment":"The caption states 'The shaded areas indicate the confidence interval,' but the text does not describe how the confidence interval is computed or over how many generations.","section":"Figure 4"},{"comment":"The limitation statement that 'visual discrepancies in the generated outputs' exist should be reconciled with the abstract's 'negligible degradation' wording, since the two statements are in tension.","section":"Section 5 (Limitation)"}],"recommendation":"major_revision","confidential_remarks":"The empirical core of the paper is solid and the method is likely to be useful, but the headline quality claim is currently overstated relative to the per-dimension results, and the context-agnostic scheduling assumption needs explicit validation. I did not find circularity: VBench is an external benchmark, and the fitted thresholds/rates are not fed back into the method in a way that would make the comparison circular. The per-model manual tuning burden should be disclosed more prominently, perhaps in the abstract's 'training-free' phrasing, since 'training-free' does not mean 'tuning-free.' The paper is within scope for a top-tier venue and the issues are fixable with additional analyses and a careful rewriting of the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a serious look. The core idea is real: instead of reducing the input H symmetrically like ToMe, AsymRnR reduces Q and K/V independently after projection, with per-feature thresholds and a matching cache that skips redundant BSM computations. That is a genuine departure from prior token merging for diffusion, and it delivers consistent 1.1–1.3x speedups across five current video DiTs, including a distilled 6-step model and stacked with PAB caching. The paper also does a fair job situating itself against ToMe and explains why ToMe fails on 3D-ROPE models. Code is released, and the authors do list honest limitations in Section 5.\n\nThe soft spots are mainly about claims versus evidence. The abstract says 'negligible degradation,' but the per-dimension VBench table tells a more nuanced story. On CogVideoX-2B, the default configuration drops Multiple Objects by 0.10 (0.6502 to 0.5480) and Color by 0.05. Those are not negligible for a user who cares about counting objects or color fidelity. The aggregate VBench score barely moves because it averages over many dimensions, so the reader has to dig into Table 11 to see what is actually happening. There are also no error bars, no multiple seeds, and no significance tests anywhere in the evaluation. That matters for a claim whose entire force is 'quality is preserved.' The limitation section admits visual discrepancies, which is honest, but the abstract does not carry that nuance.\n\nTwo smaller issues. First, the theoretical foundation (Corollary 3.1) is a consistency result for a k-NN KL estimator, not a proof that minimizing KL between token distributions preserves attention outputs. The paper uses it as motivation, which is fine, but the word 'theoretical foundations' in the abstract oversells it. Second, the reduction thresholds and rates are manually tuned per model by visual inspection (Appendix B), so 'model-agnostic' should be qualified as 'model-agnostic after a short per-model tuning session.' The authors do acknowledge this in Limitations, but it is worth stating more prominently.\n\nThe stress-test concern about per-dimension degradation is valid and lands. It does not sink the method; the speedups are measurable and the method composes with other acceleration techniques. But the authors should temper the abstract, report per-dimension scores prominently, and add at least a few seeds or error bars. If those are addressed in revision, this is a useful contribution for anyone working on efficient video generation.\n\nReading group: yes, mainly to discuss the gap between aggregate metrics and per-dimension quality. Would cite: yes, if I work on video DiT acceleration. Recommended verdict: accept with major revisions; the empirical core is sound, but the claims need to match the data.","headline":"Practical token-reduction speedup for video DiTs with a genuine asymmetric Q/KV twist, but the 'negligible degradation' headline is not supported by the paper's own per-dimension VBench numbers.","tokens_in":22433,"tokens_out":1642,"would_cite":true,"duration_ms":18174,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A training-free, model-agnostic method, AsymRnR, accelerates video diffusion transformers up to 1.3x by asymmetrically reducing redundant Q/K/V tokens, with negligible or even positive quality effects.","keywords":["video diffusion transformers","token reduction","attention acceleration","training-free acceleration","asymmetric reduction","matching cache","reduction scheduling","text-to-video generation"],"falsifier":"Generate videos from prompts whose token-similarity statistics differ systematically from the precomputed schedule, such as fast camera motion or many small independently moving objects, and measure the VBench or LPIPS gap between AsymRnR and the unmodified baseline; if the gap substantially exceeds the roughly 0.01 the paper reports across such content, the context-agnostic assumption fails. A more direct check is to compute the actual KL divergence (using the paper's nearest-neighbor estimator) between original and reduced Q/K/V sequences on held-out prompts and verify that the schedule's reductions are the ones that minimize that divergence.","tokens_in":21496,"feed_emoji":"⚡","tokens_out":4815,"duration_ms":42479,"temperature":0.7,"pith_summary":"The paper tries to establish that the attention layers of video diffusion transformers are full of redundant tokens whose redundancy differs between query tokens and key/value tokens, between early and deep blocks, and between early and late denoising steps. Exploiting this, it proposes AsymRnR, which reduces only the redundant tokens, mainly key/value tokens in high-similarity blocks and timesteps, before self-attention and restores the dropped tokens afterward so the pretrained network is barely disturbed. If the claim holds, video DiTs like CogVideoX, Mochi-1, HunyuanVideo, and a 6-step distilled FastVideo can run 1.1x to 1.3x faster with VBench scores within roughly 0.01 of baseline and sometimes slightly better, without any training. This matters because video generation cost is the main barrier to practical use, and the method plugs into existing models.","feed_headline":"Train-free token pruning speeds video diffusion up to 1.3x","feed_subtitle":"AsymRnR cuts key-value tokens that carry little information, saving compute without retraining or quality loss.","key_machinery":"The load-bearing object is the asymmetric reduction-restoration operator AsymRnR, which applies bipartite soft matching separately to Q and K/V, discards the matched redundant tokens, runs self-attention on the shortened sequence, and then restores Q to its original length by replicating each reduced token according to its match. Supporting it are three mechanisms: a reduction schedule that thresholds a precomputed similarity map S(A,t,b) per feature type, denoising timestep, and block; a matching cache that reuses the bipartite matching result for s consecutive denoising steps, cutting matching cost by a factor of 1/s; and the negative Euclidean distance as the similarity metric, which is compatible with the KL-divergence estimator of Corollary 3.1.","core_discovery":"The central claim is that matching-based token reduction can be made safe for video diffusion transformers by treating Q and K/V asymmetrically: queries act as the sensitive questioner and must be preserved or restored, while keys and values carry redundant information that can be reduced more aggressively. Combined with a reduction schedule that concentrates reduction in the blocks and timesteps where token similarity is high, and a matching cache that reuses the similarity matching across adjacent denoising steps, the method achieves 1.10x to 1.30x speedups on state-of-the-art video DiTs. The paper also grounds the design in a distributional argument: reduction should minimize the KL divergence between the reduced and original token distributions, and a nearest-neighbor Monte Carlo estimator justifies using (negative) Euclidean distance and direct token discarding rather than cosine similarity and mean-based merging.","pith_inferences":["If the context-agnostic similarity-pattern assumption holds, the same precomputed schedule could be reused across prompts with zero per-prompt profiling; a natural stress test is to deploy AsymRnR on out-of-distribution prompts such as fast camera motion or scenes with many small independently moving objects and check whether the claimed speed-quality tradeoff persists.","The KL-divergence framing suggests a stronger criterion than thresholding: one could directly minimize the estimated nearest-neighbor divergence per layer, yielding a content-adaptive per-token budget rather than fixed thresholds.","The Q-versus-K/V asymmetry may carry over to cross-attention and autoregressive attention in video models, where the questioner role also differs, though the paper only treats self-attention.","The matching cache exploits temporal smoothness of token matches; the same lazy-update idea could be applied to the reduction schedule itself, refreshing the schedule as the video content evolves rather than fixing it once."],"forward_implications":["Applying AsymRnR to CogVideoX-2B, CogVideoX-5B, Mochi-1, HunyuanVideo, and the 6-step distilled FastVideo-Hunyuan yields 1.10x to 1.30x latency reductions with VBench scores within roughly 0.01 of baseline, and slightly higher than baseline on HunyuanVideo and Mochi-1.","AsymRnR is orthogonal to feature caching: stacked on top of PAB on HunyuanVideo it achieves a combined 1.71x speedup with negligible quality degradation.","AsymRnR is training-free and model-agnostic, and it also works on UNet-based video diffusion models, giving a 1.20x speedup on AnimateDiff with no perceptible quality loss.","Using negative Euclidean distance and direct token discarding instead of cosine similarity and mean-based merging improves both output quality and latency, with the quality benefit most visible when reducing value tokens.","Larger models tolerate the reduction better than smaller ones, and the method composes with step-distilled few-step samplers without further tuning."],"supporting_citations":[{"why":"Supplies bipartite soft matching, the token-merging primitive that AsymRnR adapts and makes asymmetric.","marker":"(Bolya et al., 2023)"},{"why":"The ToMe token-reduction baseline for diffusion models that the paper compares against and critiques for symmetric, uniform reduction.","marker":"(Bolya & Hoffman, 2023)"},{"why":"Provides the k-nearest-neighbor KL-divergence estimator that underlies Corollary 3.1 and motivates Euclidean-distance matching.","marker":"(Wang et al., 2009)"},{"why":"The VBench suite used for the primary quantitative quality evaluation across 16 video-generation dimensions.","marker":"(Huang et al., 2024)"},{"why":"CogVideoX, the main DiT testbed where AsymRnR is compared against ToMe and ablations are run.","marker":"(Yang et al., 2024)"},{"why":"HunyuanVideo, a state-of-the-art video DiT where AsymRnR achieves the largest relative speedup.","marker":"(Team, 2024c)"},{"why":"PAB feature caching, the method AsymRnR is stacked with to demonstrate orthogonality and a 1.71x combined speedup.","marker":"(Zhao et al., 2024)"}],"fun_headline_variants":["AsymRnR: 1.3x speedup for video diffusion, training-free","Token pruning accelerates video diffusion 1.3x, no retrain","Asymmetric reduction: key to faster video diffusion","Drop redundant tokens, speed up video generation 30%","AsymRnR: Clip token fat for 1.3x faster video diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme rests on two unproven premises: that minimizing KL divergence between the reduced and original token distributions is the right objective for preserving attention outputs, and that per-token similarity patterns are context-agnostic, so a schedule estimated on arbitrary prompts transfers to any new prompt.","fun_headline_variants_meta":{"raw":{"variants":["AsymRnR: 1.3x speedup for video diffusion, training-free","Token pruning accelerates video diffusion 1.3x, no retrain","Asymmetric reduction: key to faster video diffusion","Drop redundant tokens, speed up video generation 30%","AsymRnR: Clip token fat for 1.3x faster video diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000809,"raw_usage":{"total_tokens":3520,"prompt_tokens":882,"completion_tokens":2638,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":2553}},"tokens_in":498,"tokens_out":2638,"duration_ms":18033,"temperature":1.0,"reasoning_tokens":2553,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:38:46.456712+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate videos from prompts whose token-similarity statistics differ systematically from the precomputed schedule, such as fast camera motion or many small independently moving objects, and measure the VBench or LPIPS gap between AsymRnR and the unmodified baseline; if the gap substantially exceeds the roughly 0.01 the paper reports across such content, the context-agnostic assumption fails. A more direct check is to compute the actual KL divergence (using the paper's nearest-neighbor estimator) between original and reduced Q/K/V sequences on held-out prompts and verify that the schedule's reductions are the ones that minimize that divergence.","supporting_citations":[{"cited_title":"R., and Verd \\' u , S","cited_arxiv_id":null,"evidence_quote":"Provides the k-nearest-neighbor KL-divergence estimator that underlies Corollary 3.1 and motivates Euclidean-distance matching."}],"review_version":1}