acceptodds
Under review as a conference paper at ICLR 2027

DeCoPrune: Efficient KV-Cache Pruning for Autoregressive Video Diffusion via Denoising Consistency

11 views

Abstract

Autoregressive video diffusion naturally supports streaming generation and interactive control, but its KV cache grows continuously with the generated history. Existing compression strategies either discard history using fixed windows or select tokens through local attention and similarity signals, which do not directly measure whether the current chunk contributes information beyond the retained context. We introduce DeCoPrune, a training-free method that treats cache compression as a denoising-consistency problem. We find empirically that denoising difficulty provides a useful proxy for a token's value in long-term retention: tokens with larger step-to-final discrepancies tend to carry visual evidence that is less predictable from the retained context. Based on this finding, DeCoPrune measures each current-chunk token's denoising difficulty using the discrepancy between its intermediate clean prediction and final denoised value, retaining high-discrepancy tokens in the long-term cache while pruning those with low discrepancy. To evaluate information retention, we introduce CMBench, comprising 58 approximately one-minute generated or real-world context episodes and 116 Reappear or Revisit continuation tasks that require recalling events or objects shown earlier in the context. Experiments with LingBot World v2 show that DeCoPrune achieves a DINO score of 0.6701 on a 0–1 scale with an 85.43% reduction in cumulative historical KV token counts and a 4.14× continuation-generation speedup over FullKV. Its head-specialized variant further reaches 0.6783 at an 86.19% pruning ratio, approaching the 0.6803 score of FullKV and exceeding the evaluated compression baselines at similar budgets. These results indicate that denoising consistency can serve as a model-intrinsic signal for retaining long-range information while reducing autoregressive inference cost.

Primary area:
generative models
Keywords:
ar video generation, kv cache prune, generation consistency

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

This paper introduces DeCoPrune, a training-free KV-cache pruning method for autoregressive video diffusion that removes tokens whose intermediate clean predictions closely match final outputs, reusing denoising predictions at negligible cost. It also introduces CMBench, comprising 116 reference-grounded Reappear/Revisit tasks over minute-scale contexts. On LingBot World v2, DeCoPrune achieves near-FullKV recall (0.678 vs. 0.680 DINO) at ~86% pruning and a 4.1× continuation speedup, outperforming evaluated compression baselines at comparable budgets.

Critical strengths
  1. The paper is easy to read and follow.

  2. Novel, model-intrinsic criterion. Using step-to-final denoising discrepancy to identify tokens redundant given the retained context is an original and elegant idea. It reuses existing predictions without auxiliary models, retrieval indices, motion estimators, or additional scoring overhead.

  3. Strong qualitative and quantitative results.

  4. Well-motivated benchmark. CMBench directly measures contextual recall through reference-grounded evaluation. The paper demonstrates its value by showing that standard quality metrics remain saturated even as contextual recall deteriorates.

  5. Convincing ablations. At matched budgets, DeCoPrune (0.670) clearly outperforms both Random (0.609) and the reverse criterion (0.451), supporting the proposed selection direction. The ablations also cleanly separate the contribution of RoPE re-indexing.

Critical weaknesses
  1. Limited evidence of generalization. The main results rely on a single backbone, LingBot World v2, while validation on LongLive covers only 18 cases with 10-second contexts.

  2. Benchmark and evaluation limitations. CMBench is largely synthetic: 50 of 58 episodes are generated by another video model, with only 14 real-world tasks, raising concerns about distribution bias. The OWL-ViT detection and max-over-frames DINO protocol also has limitations: a single matching frame can yield a high score, while detection failure forces the score to zero.

  3. Unbounded memory and future relevance. The retained cache continues to grow with rollout length, and the criterion does not explicitly account for future relevance: a rare detail that is easy to denoise now may be discarded despite being needed later. Although both limitations are acknowledged, discussing bounded-memory or task-aware extensions would strengthen the paper.

Questions for authors

The paper presents an original and well-ablated training-free pruning criterion with strong results, but its reliance on a single backbone and a mostly synthetic self-constructed benchmark leaves generalization insufficiently established. Please refer to weaknesses above.

4 comments
AuthorsOpenReview ↗

Response to Reviewer viwi (1/2)

We are grateful for the reviewer’s positive assessment and helpful comments. We respond to the concerns below.

W1 / Q1 — Generalization beyond a single backbone.


We use LingBot World v2 because it is the only backbone we tested that actually recalls content from a one-minute context. Even with FullKV, LongLive, Self-Forcing and Causal Forcing score about 0.26 DINO, compared with about 0.68 for LingBot (Appendix C). With such a weak unpruned reference, the effect of pruning cannot be isolated. The denoising signal behind our criterion holds on all four backbones: in Fig. 2(b), each shows lower discrepancy when the relevant content is already in the history.

Following your suggestion, we expanded the 10-second LongLive evaluation from 18 cases to all 116 CMBench tasks. Each task uses the clip that contains its annotated reference frame, with the original continuation prompt, target, reference box and scoring. We widened the budgets of the four LongLive cache policies so that all methods run at a comparable pruning ratio (62–65%). At 62.38% pruning, DeCoPrune scores 0.5460 DINO versus 0.5486 for FullKV, and it scores 0.0558–0.0874 higher than all four policies, with every paired 95% CI excluding zero (Table R1). TempDiff is close to DeCoPrune here because, with only a 10-second context, the target lies a few chunks back and any budget-matched selector tends to keep it; on minute-scale LingBot contexts, DeCoPrune outperforms TempDiff by 0.0537 (paired 95% CI [+0.0181, +0.0902]).

Table R1. LongLive, CMBench (116 tasks, seed 2), 10 s context. All compression methods run at a comparable PR (62–65%). Δ = DeCoPrune − comparator (paired bootstrap 95% CI).

MethodPRDINO [95% CI]Paired Δ [95% CI]
FullKV0.00%0.5486 [0.5005, 0.5964]−0.0026 [−0.0507, +0.0458]
DeCoPrune62.38%0.5460 [0.5053, 0.5857]—
TempDiff62.39%0.5452 [0.5048, 0.5851]+0.0007 [−0.0220, +0.0235]
DummyForcing64.85%0.4902 [0.4586, 0.5220]+0.0558 [+0.0226, +0.0909]
Streaming62.41%0.4853 [0.4530, 0.5188]+0.0607 [+0.0295, +0.0937]
LongLive-RAG62.41%0.4728 [0.4386, 0.5075]+0.0732 [+0.0392, +0.1082]
ForcingKV62.77%0.4585 [0.4264, 0.4911]+0.0874 [+0.0547, +0.1213]
AuthorsOpenReview ↗

Response to Reviewer viwi (2/2)


W2 — Synthetic-heavy benchmark and max-over-frames scoring.


Synthesis broadens CMBench's coverage, and most synthetic videos depict real-world settings. Generation lets us include stylized domains such as anime and game worlds and control exactly when each target appears; 38 of the 50 synthetic videos (76%) are photorealistic clips of real-world settings. The real subset checks that conclusions transfer: DeCoPrune ranks first among the pruning methods on both real and synthetic footage, and the two subsets agree on the ordering of 14 of 15 method pairs (Kendall's τ = 0.87, Table 6 in Appendix D).

The comparison holds when several frames must score well. We re-scored all methods with the mean of the top-3 and the top-5 of the 10 sampled frames (Table R2). With top-3, the ranking of all 12 methods is identical to the max ranking. With top-5, DeCoPrune ranks first among all methods, including FullKV, and its paired CIs against TempDiff, the context-free control and Random still exclude zero. We also manually reviewed all cases and found none in which OWL-ViT missed a successfully generated target, so a zero score corresponds to a genuine recall failure.

Table R2. DINO under three frame aggregations on 116 CMBench tasks (LingBot World v2). Δ = DeCoPrune − comparator (paired bootstrap 95% CI).

MethodMaxTop-3Top-5Δ top-3 [95% CI]Δ top-5 [95% CI]
FullKV0.69350.65830.6080−0.0069 [−0.0436, +0.0310]+0.0027 [−0.0342, +0.0402]
DeCoPrune0.68650.65140.6107——
DensityKV0.67810.64270.6064+0.0087 [−0.0208, +0.0398]+0.0043 [−0.0250, +0.0359]
TempDiff0.63280.59340.5503+0.0580 [+0.0226, +0.0946]+0.0604 [+0.0262, +0.0961]
Context-free0.62670.58830.5548+0.0631 [+0.0362, +0.0917]+0.0559 [+0.0294, +0.0842]
Random0.59760.56220.5188+0.0892 [+0.0558, +0.1233]+0.0920 [+0.0605, +0.1251]
Future Forcing0.59690.55820.5242+0.0932 [+0.0637, +0.1243]+0.0866 [+0.0574, +0.1175]
WorldKV0.55420.52090.4864+0.1305 [+0.0878, +0.1751]+0.1243 [+0.0817, +0.1685]
SnapKV-style0.55330.52050.4931+0.1309 [+0.0959, +0.1672]+0.1176 [+0.0828, +0.1537]
ForcingKV0.54610.51270.4786+0.1387 [+0.0921, +0.1884]+0.1321 [+0.0877, +0.1798]
DummyForcing0.45380.42400.3960+0.2274 [+0.1876, +0.2679]+0.2147 [+0.1768, +0.2543]
Streaming0.44880.41960.3968+0.2318 [+0.1925, +0.2729]+0.2139 [+0.1765, +0.2533]

W3 — Unbounded memory and future relevance.


Rare details tend to be retained rather than discarded. When the context provides little evidence about a rare detail, the model is uncertain about its appearance, so its tokens have high denoising discrepancy and are kept. A detail is easy to denoise mainly when the cache already contains it.

We agree that the current setting has no fixed memory budget. DeCoPrune reduces the KV retained per chunk, but the cache still grows with rollout length. Because every token has a score, a bounded variant is direct: keep a short recent window and evict the lowest-scoring older tokens whenever the cache reaches a fixed budget. Where task cues are available, such as a requested viewpoint in a camera-conditioned world model, the eviction policy can protect the tokens likely to be needed next, in the spirit of WorldMem (arXiv:2504.12369) and Context-as-Memory (arXiv:2506.03141). We will state this limitation and discuss both extensions in the revision.

Thank you for the rebuttal, which addressed my concerns and I belive most of the other reviewer's concerns. I have decided to maintain my score.

Thank you for taking the time to read our rebuttal and follow up. We’re glad it addressed your concerns and appreciate your constructive feedback.

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

This paper proposes DeCoPrune, a training-free KV-cache pruning method for autoregressive video diffusion. For each token, it computes the discrepancy between the x^0\hat{x}_0 prediction at an intermediate denoising step and the final output, keeps tokens with high discrepancy, and removes the rest from the cache, while retaining the sink chunk and the most recent chunk. It is combined with RoPE re-indexing and a static/dynamic head partition (DeCoPrune-HS). The paper also introduces CMBench, a benchmark of 116 tasks that measures whether objects from a one-minute context reappear correctly in the continuation, and evaluates on LingBot World v2 against Streaming, DummyForcing, ForcingKV and TempDiff.

Critical strengths
  • The retention signal is simple and cheap: it reuses the x^0\hat{x}_0 prediction that the sampler already computes, with no extra model or training.

  • The pruning physically removes tokens from the cache, and the reported FPS confirms the speedup is realized in practice.

  • CMBench addresses a real gap. Fig. 6b shows that VBench consistency metrics stay flat while object recall drops sharply as the pruning ratio increases, which is a good argument for a recall-based benchmark.

Critical weaknesses
  • The differences in Table 1 are marginal and no std is reported. The gap between DeCoPrune-HS and FullKV is 0.002 and between DeCoPrune-HS and DeCoPrune is 0.008, from a three-seed mean over 116 tasks. Without variance these numbers cannot be distinguished from noise. The VBench quality metrics in the same table are also essentially identical across all methods.

  • I do not understand the relation between Table 1 and Table 2. Section 4.2 states that all methods in Table 1 use RoPE re-indexing, so the "w/ re-index" column of Table 2 should reproduce Table 1? But DeCoPrune-HS is 0.7156 in Table 2 and 0.6783 in Table 1, and every other method also changes by 0.01–0.06, while FullKV is unchanged. Also in Table 2 the pruned cache is also above FullKV, in which case FullKV is not the upper bound the paper treats it as.

  • The baselines comparison is limited. The content-aware policies cited in related work (Focused Forcing, Future Forcing, DensityKV, WorldKV) are not evaluated. And the introduction positions DeCoPrune against local-similarity pruning, but the main competitor TempDiff is the authors' own construction: L352-354 "TempDiff is a temporal-difference baseline inspired by temporal-redundancy-aware token reduction".

Questions for authors

Please refer to Weaknesses section.

5 comments
AuthorsOpenReview ↗

Response to Reviewer ovfR (1/2)

We appreciate the reviewer’s careful reading and constructive feedback. Our responses follow below.

W1 — Marginal differences, missing variance, and flat VBench.


We added bootstrap 95% CIs and paired tests on the 116 CMBench tasks (Table R1). Compared with FullKV, DeCoPrune prunes 86.60% of historical KV, runs 4.35× faster, and loses only 0.0070 DINO (95% CI [−0.0450, +0.0318]). A small gap to FullKV is exactly what a pruning method aims for. Compared with the other compression methods, DeCoPrune scores higher than all of them, by +0.0373 (Focused Forcing) to +0.2377 (Streaming). The paired CIs exclude zero for all of them except DensityKV, a concurrent work (+0.0084), which runs at only 0.48× FullKV's speed.

DeCoPrune-HS is a head-specialized variant that uses the same criterion. We do not claim that it beats DeCoPrune; the 0.002 and 0.008 gaps in Table 1 are descriptive differences between three-seed means, and we will make this clear.

VBench is nearly flat because it measures general video quality, which barely changes when the same frozen generator runs with different caches. In Fig. 6(b), VBench consistency hardly moves as PR increases from 60% to 95%, while CMBench DINO falls from about 0.78 to 0.52. CMBench directly measures the recall that pruning must preserve.

Table R1. Results on 116 CMBench tasks (LingBot World v2). Δ = DeCoPrune − comparator on the same tasks. Bootstrap 95% CIs; paired Wilcoxon signed-rank p-values. FPS measures autoregressive continuation on four H200 GPUs; speedup is relative to FullKV.

MethodPRDINO [95% CI]Paired Δ [95% CI]Wilcoxon pFPSSpeedup
FullKV0.00%0.6935 [0.6522, 0.7331]−0.0070 [−0.0450, +0.0318]0.1381.4301.000×
DeCoPrune86.60%0.6865 [0.6525, 0.7196]——6.2274.354×
DensityKV85.48%0.6781 [0.6436, 0.7113]+0.0084 [−0.0221, +0.0400]0.5680.6890.482×
Focused Forcing85.50%0.6491 [0.6082, 0.6875]+0.0373 [+0.0045, +0.0726]0.2513.5142.457×
TempDiff86.35%0.6328 [0.5960, 0.6694]+0.0537 [+0.0181, +0.0902]0.0316.4354.500×
Random86.78%0.5976 [0.5614, 0.6346]+0.0888 [+0.0551, +0.1234]1e-66.3504.440×
Future Forcing85.65%0.5969 [0.5614, 0.6323]+0.0895 [+0.0601, +0.1208]9e-84.0432.827×
WorldKV85.29%0.5542 [0.5143, 0.5935]+0.1322 [+0.0896, +0.1773]2e-75.5623.889×
SnapKV-style selector85.65%0.5533 [0.5175, 0.5893]+0.1332 [+0.0973, +0.1703]2e-105.7774.040×
ForcingKV80.63%0.5461 [0.5047, 0.5876]+0.1404 [+0.0914, +0.1913]1e-65.0893.559×
DummyForcing98.12%0.4538 [0.4209, 0.4876]+0.2326 [+0.1917, +0.2748]7e-177.6075.319×
Streaming96.78%0.4488 [0.4164, 0.4807]+0.2377 [+0.1966, +0.2804]9e-189.3546.541×
AuthorsOpenReview ↗

Response to Reviewer ovfR (2/2)


W2 — Relationship between Tables 1 and 2; FullKV as an upper bound.


Thank you for catching this. Table 2 is a diagnostic study of re-indexing on 39 tasks with seed 2, whereas Table 1 covers all 116 tasks with three seeds, so its "w/ re-index" column does not reproduce Table 1. The unchanged FullKV entry is a reporting error: it was copied from Table 1, and the correct value on the 39 tasks is 0.6714. All other entries are correct. We will fix this entry (and the same number in the Fig. 7 caption) and state the protocol of Table 2 explicitly.

The paper treats FullKV as the uncompressed reference rather than an upper bound (L408–409). A pruned cache can score higher on a subset for two reasons. Removing redundant history leaves fewer irrelevant tokens competing for attention with the needed evidence. Retrieval from a long uncompressed history is also sensitive to temporal position: re-indexing alone changes FullKV by +0.2329 on this subset. On all 116 tasks, DeCoPrune and FullKV are not significantly different (Table R1).


W3 — Content-aware baselines; TempDiff as the authors' construction.


We added all four cited content-aware methods and an attention-based selector (Table R1). At similar pruning ratios, DeCoPrune scores 15.0%, 23.9% and 24.1% higher DINO than Future Forcing, WorldKV and the SnapKV-style selector, and runs faster than each. It scores 5.8% higher than our Focused Forcing adaptation at 1.77× its throughput, and higher than DensityKV (0.6865 vs. 0.6781) at 9.0× its throughput.

TempDiff is a controlled test of local-similarity pruning. EVEREST and FrameFusion target masked autoencoding and VLM token reduction rather than AR-diffusion KV caches, so they must be adapted. TempDiff keeps their shared rule, ranking patches by how much they change from the previous frame, and uses the same windows, budget and re-indexing as DeCoPrune. At similar pruning ratios, DeCoPrune improves DINO over TempDiff by 0.0537 (+8.5%, paired 95% CI [+0.0181, +0.0902]).

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer ovfR,

Thank you again for your detailed and constructive review. As the discussion period closes soon, we would like to kindly check whether our response has addressed your concerns.

If any point remains unclear or insufficient, we would be very grateful for your further feedback and are happy to provide additional clarification or results before the deadline. We look forward to hearing from you.

Best regards,

The Authors

Reviewer_9oYyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

DeCoPrune is a training-free KV-cache pruning rule for autoregressive video diffusion. It keeps a token in the long-term cache if its clean prediction at an intermediate denoising step differs strongly from the final output, on the premise that tokens already explained by the cache converge early. Observed contexts are handled by re-noising each chunk. The paper also introduces CMBench, 116 recall tasks over one-minute contexts scored by DINOv2 similarity to the referenced object. On LingBot World v2, DeCoPrune reaches 0.670 DINO (FullKV: 0.680) at 85% cumulative pruning with a 4.14× continuation speedup, outperforming Streaming, DummyForcing, ForcingKV, and a temporal-difference baseline.

Critical strengths

The score needs no auxiliary predictor, retrieval index, or training, and in the online case it reuses predictions that are already computed. The reversed-criterion ablation is the strongest result in the paper. At matched pruning (≈85%), keeping low-discrepancy tokens drops DINO from 0.670 to 0.451, and keeping 81% of tokens in reverse order (0.562) is still worse than keeping the top 15% by discrepancy. Whatever the score captures, it ranks history tokens in a useful way, and that is what would make the method practical.

Fig. 6(b) shows VBench subject and background consistency staying flat while recall collapses as pruning increases. A benchmark with reference-grounded targets, and Revisit tasks that put the target far outside the recent window, fills a real gap in a crowded area where most papers report only quality metrics.

Critical weaknesses
  1. The evaluated setting differs from the claimed one. CMBench compresses an observed one-minute video before a short continuation, so most pruning uses the prefix variant, which requires an extra forward pass per chunk, or is it not?

  2. The core hypothesis and comparisons are under-tested. Fig. 2(b) shows a chunk-level effect, but high token discrepancy could reflect intrinsic texture or motion difficulty rather than redundancy with the cache, and a context-free control is missing. No confidence intervals are given, the headline gaps are small, and Random scores 0.609 in Table 3 but 0.672 in Table 2 (close to FullKV) without explanation. Attention-based selection and the cited content-aware video KV methods are not compared, and the main results rely on a single backbone.

Questions for authors

Does the prefix variant lead to an extra forward pass per chunk. If not, how does it work?

What is the "diagnostic evaluation" in Table 2, and why does Random differ so much from Table 3? Please add Random to Table 1 with confidence intervals.

How does scoring by discrepancy computed without the history cache compare at matched pruning?

Can you compare with an attention-based selector and at least one cited content-aware method (e.g., Focused Forcing) at matched PR? Also, will CMBench (prompts, references, boxes, and the real clips, with their licenses) be released?

7 comments
AuthorsOpenReview ↗

Response to Reviewer 9oYy (1/3)

We thank the reviewer for the encouraging assessment and thoughtful suggestions. We address each point below.

W1 / Q1 — Does the prefix variant need an extra forward pass?


Yes. For an observed context, the prefix variant adds one denoising pass per chunk. Observed footage has no denoising trajectory, so we re-noise each context chunk to τ* and denoise it once, conditioned on the compressed cache. Generated chunks already have a denoising trajectory, so DeCoPrune scores them without any extra model call. The criterion (Eq. 1) is the same in both cases.

CMBench uses an observed prefix so that every method must retrieve the same evidence from the same history, which isolates long-context recall. In interactive generation and world exploration, most of the history is generated by the model itself and is scored at no extra cost.

Even with the prefix passes, DeCoPrune is 3.25× faster than FullKV end to end (Table R1). Its context phase is also faster than FullKV's plain prefill (142 s vs. 251 s), because later context chunks attend to a much smaller cache.

Table R1. Wall-clock time per task (median over 30 tasks on one GPU; 60 s observed context + 30 s generation).

MethodContext phase (incl. prefix scoring)GenerationEnd-to-endE2E speedup
FullKV251 s859 s1110 s1.00×
DeCoPrune142 s199 s341 s3.25×

W2 We address each point of W2 in turn: (a) the hypothesis and the context-free control, (b) confidence intervals, (c) Random in Tables 2 and 3, (d) additional baselines, (e) the single backbone, and (f) the CMBench release.


(a) Texture and motion vs. redundancy; context-free control (also Q3).

Fig. 2(b) already holds texture and motion fixed. For LingBot World v2, both runs rotate the camera in the same way and reveal the same content; the only difference is whether that content appeared earlier in the context. The revealed view and the camera motion are identical, so intrinsic texture and motion difficulty are the same in both runs. Discrepancy is consistently lower when the content was seen before, so the score reflects what the cache already contains. Table 3 also confirms the direction of the criterion: at comparable pruning ratios (85.31–85.43%), keeping high-discrepancy tokens scores 0.6701 DINO, random selection 0.6091, and keeping low-discrepancy tokens 0.4512.

For Q3, removing the history cache from the score lowers DINO from 0.6865 to 0.6267 at a nearly identical pruning ratio (Table R2). The control computes the same discrepancy at the same denoising step without the history cache, and keeps the same number of tokens per chunk as DeCoPrune. The context-free score improves on Random by only 0.0291, whereas conditioning on the cache improves it by 0.0888. If discrepancy mainly reflected texture or motion, the context-free score would capture it and perform close to DeCoPrune.

Table R2. Retention signals on 116 CMBench tasks (LingBot World v2). Δ = DeCoPrune − comparator on the same tasks.

Retention signalPRDINO [95% CI]Paired Δ [95% CI]Wilcoxon p
Cache-conditioned discrepancy (DeCoPrune)86.60%0.6865 [0.6525, 0.7196]——
Context-free discrepancy86.75%0.6267 [0.5896, 0.6638]+0.0598 [+0.0321, +0.0894]1e-4
Random86.78%0.5976 [0.5614, 0.6346]+0.0888 [+0.0551, +0.1234]1e-6
AuthorsOpenReview ↗

Response to Reviewer 9oYy (2/3)

W2 (continued).


(b) Confidence intervals and small gaps.

We re-ran all methods on all 116 tasks with a shared seed and report bootstrap 95% CIs and paired tests on the same tasks (Table R3). DeCoPrune is within 0.0070 DINO of FullKV (95% CI [−0.0450, +0.0318]) while pruning 86.60% of historical KV and running 4.35× faster. It scores higher than every other compression method, and the paired CIs exclude zero for all of them except DensityKV, a concurrent work (+0.0084). DensityKV runs at only 0.48× FullKV's speed.

Table R3. Results on 116 CMBench tasks (LingBot World v2). Δ = DeCoPrune − comparator on the same tasks. Bootstrap 95% CIs; paired Wilcoxon signed-rank p-values. FPS measures autoregressive continuation on four H200 GPUs; speedup is relative to FullKV.

MethodPRDINO [95% CI]Paired Δ [95% CI]Wilcoxon pFPSSpeedup
FullKV0.00%0.6935 [0.6522, 0.7331]−0.0070 [−0.0450, +0.0318]0.1381.4301.000×
DeCoPrune86.60%0.6865 [0.6525, 0.7196]——6.2274.354×
DensityKV85.48%0.6781 [0.6436, 0.7113]+0.0084 [−0.0221, +0.0400]0.5680.6890.482×
Focused Forcing85.50%0.6491 [0.6082, 0.6875]+0.0373 [+0.0045, +0.0726]0.2513.5142.457×
TempDiff86.35%0.6328 [0.5960, 0.6694]+0.0537 [+0.0181, +0.0902]0.0316.4354.500×
Random86.78%0.5976 [0.5614, 0.6346]+0.0888 [+0.0551, +0.1234]1e-66.3504.440×
Future Forcing85.65%0.5969 [0.5614, 0.6323]+0.0895 [+0.0601, +0.1208]9e-84.0432.827×
WorldKV85.29%0.5542 [0.5143, 0.5935]+0.1322 [+0.0896, +0.1773]2e-75.5623.889×
SnapKV-style selector85.65%0.5533 [0.5175, 0.5893]+0.1332 [+0.0973, +0.1703]2e-105.7774.040×
ForcingKV80.63%0.5461 [0.5047, 0.5876]+0.1404 [+0.0914, +0.1913]1e-65.0893.559×
DummyForcing98.12%0.4538 [0.4209, 0.4876]+0.2326 [+0.1917, +0.2748]7e-177.6075.319×
Streaming96.78%0.4488 [0.4164, 0.4807]+0.2377 [+0.1966, +0.2804]9e-189.3546.541×

(c) Random in Tables 2 and 3 (also Q2).

Tables 2 and 3 use different task sets. Table 2 is a diagnostic study of RoPE re-indexing on 39 tasks with seed 2; Table 3 evaluates token selection on all 116 tasks. We will state each table's protocol, and add Random with its CI to Table 1 (Table R3: 0.5976 [0.5614, 0.6346]).

AuthorsOpenReview ↗

Response to Reviewer 9oYy (3/3)

W2 (continued).


(d) Attention-based and content-aware baselines (also Q4).

We added an attention-based (SnapKV-style) selector and all four cited content-aware methods at similar pruning ratios (Table R3). DeCoPrune scores 24.1%, 15.0% and 23.9% higher DINO than the SnapKV-style selector, Future Forcing and WorldKV, and runs faster than each. Against our Focused Forcing adaptation on LingBot World v2, it scores 5.8% higher at 1.77× the throughput (6.227 vs. 3.514 FPS). Against DensityKV, it scores 0.6865 vs. 0.6781 at 9.0× the throughput (6.227 vs. 0.689 FPS).

(e) Single backbone.

We use LingBot World v2 because it is the only backbone we tested that actually recalls content from a one-minute context. With FullKV, it scores about 0.68 DINO, whereas LongLive, Self-Forcing and Causal Forcing score about 0.26 (Appendix C). With such a weak unpruned reference, the effect of pruning cannot be isolated. The denoising-consistency effect in Fig. 2(b) holds on all four backbones.

We also expanded the 10-second LongLive evaluation from 18 cases to all 116 CMBench tasks. Each task uses the clip that contains the annotated reference frame, with the continuation prompt, target and scoring unchanged. We widened the budgets of the four LongLive cache policies so that all methods run at a comparable pruning ratio (62–65%). DeCoPrune is within 0.0026 DINO of FullKV and scores 0.0558–0.0874 higher than all four policies, with every paired 95% CI excluding zero (Table R4). TempDiff is close to DeCoPrune here because, with only a 10-second context, the target lies a few chunks back and any budget-matched selector tends to keep it; on minute-scale LingBot contexts, DeCoPrune outperforms TempDiff by 0.0537 (Table R3).

Table R4. LongLive, CMBench (116 tasks, seed 2), 10 s context. All compression methods run at a comparable PR (62–65%). Δ = DeCoPrune − comparator (paired bootstrap 95% CI).

MethodPRDINO [95% CI]Paired Δ [95% CI]
FullKV0.00%0.5486 [0.5005, 0.5964]−0.0026 [−0.0507, +0.0458]
DeCoPrune62.38%0.5460 [0.5053, 0.5857]—
TempDiff62.39%0.5452 [0.5048, 0.5851]+0.0007 [−0.0220, +0.0235]
DummyForcing64.85%0.4902 [0.4586, 0.5220]+0.0558 [+0.0226, +0.0909]
Streaming62.41%0.4853 [0.4530, 0.5188]+0.0607 [+0.0295, +0.0937]
LongLive-RAG62.41%0.4728 [0.4386, 0.5075]+0.0732 [+0.0392, +0.1082]
ForcingKV62.77%0.4585 [0.4264, 0.4911]+0.0874 [+0.0547, +0.1213]

(f) CMBench release (also Q4).

Yes, we will release CMBench in full: all 58 context videos including the real clips, the prompts, reference frames and target boxes, and the generation and evaluation code. All real-world clips were recorded by the authors and contain no faces, other personal information or copyrighted material.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.