acceptodds
Under review as a conference paper at ICLR 2027

SuBitKV: Sub-Bit KV Cache Compression for Autoregressive Video Generation

20 views

Abstract

Autoregressive video generation relies on a growing key–value (KV) cache to retain information from previous frames, which raises a memory challenge as video length increases. However, retaining the full KV cache at high precision becomes impractical under limited GPU memory. Discarding past states or reducing their precision saves memory but can compromise consistency in appearance, structure, and motion over time. To address this challenge, we introduce SuBitKV, a training-free KV cache compression framework that supports storage below one bit per original KV element. Its key idea is to preserve a low-memory approximation of the retained KV cache as a persistent base, then allocate additional bits across the KV cache within the memory budget to improve reconstruction fidelity. We construct this persistent base using residual vector quantization, which jointly represents the elements of each key or value vector with a small set of compact indices rather than storing each element separately. As the cache grows, SuBitKV reduces these extra bits to stay within budget while retaining the persistent base. Experiments on three video-generation and world-model backbones show a strong memory–quality trade-off. On Wan2.2, SuBitKV achieves compression with near-BF16 VBench scores and reaches under tighter memory budgets.

Primary area:
generative models
Keywords:
V cache compression, video generation, residual vector quantization, low-bit quantization, memory-efficient inference, autoregressive video generation, sub-bit compression

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ZKtKOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a training-free VQ-like (due to residual learning) method for compressing the KV-cache used by autoregressive video generators. Since the method is vector-quantization based, unlike traditional quantization, it by definition has the ability to provide sub-bit quantization capability. Although sounds like an interesting approach, the paper fails to properly evaluate and motivate the proposed approach, barring aside the robustness and the generalizability of the proposed approach.

A summary of the approach is: a. For all the KV vectors in the cache, reconstruct them based on a persistent base + learned residuals. b. The persistent bases are heuristically adaptive in order to maintain a budget Q. c. The learned residuals are also heuristically adaptive in order to maintain the same budget Q. d. Both b. and c. are chunkwise heuristically adapted depending on the recency of the chunks. In short, it's a delicately engineered and handcrafted approach.

Critical strengths

None

Critical weaknesses
  1. The biggest critical weakness, I believe, is that the main motivation of this paper is quite cloudy. First of all, KV-caching is not really a thing for the native Video Generation (text / image conditioned or otherwise) models. This is primarily because they are all Diffusion based and not Autoregressive, unlike LLMs where it's an indispensable component. Two of the main experiments that this paper does are video generation which are not very meaningful. The HY-WorldPlay 8B experiments make sense, being a camera, action conditioned world model, but again, in this setting the chosen baselines are biased. Overall, the baselines are biased towards: "Retain all KVs, and quantize them all", which would have been fair, if SuBitKV also did the same, but as mentioned earlier in the summary, they have a heuristic soft dropping of KV entries from further past. That's why two straightforward cache-compression baselines are quite important to compare against. First, Equal-memory BF16 recent-window; and second uniformly sampled historical-chunk baseline. Without this, the complexity and the trouble of incorporating SubitKV is not justified.

  2. Highly underspecified setting of the Wan2.2 5B model, which is used quite heavily in all the experiments of the paper (also in motivation / analysis). The native generation of this model is bidirectional single chunk denoising for 31 latent / 121 RGB frames. How exactly was this native DiT transformed into your causal chunkwise generation setting of 11 latent chunk generation with 4 frames per chunk leading to 44 latent / 173 RGB frames? This setting is vastly different from the original model, and without significant post-training, similar to self-forcing, cannot lead to any reliable analysis and thus conclusions. It's imperative to specify this properly in the paper. If the model was indeed post-trained, then the claims of being a training-free method are not valid.

  3. Apart from the Wan 2.2 5B setting, just everywhere the paper is so underspecified. No details are provided in so many places. For instance Figure 2. What exactly is being computed here? What is the BF16 baseline? If BF16 baseline is just native generation, then why does it even have any temporal jumps? Are you operating on a non-post-trained-causalified DiT backbone? If yes, then all this analysis is invalid. This should be done on something like Self-Forcing (or CausVid). Figure 3: Which model's which layer's KV vectors are being visualized here? Which calibration prompts? These are so cryptic, that it's impossible to distill meaningful information. Same with Figure 5, which residual energy is being referred to here? Where is that defined?

  4. Barring aside all these problems of the paper, one of the main thing stands out is the actual SuBit compression quality is extremely bad. For instance Wan 2.25B for Q=0.2 ~0.747 bits/scalar and Q=0 ~0.532 bits/scalar, the PSNRs are 18.32 and 16.30, which is absolutely non-usable settings. For training-free methods the VBench is not really a correct metric, because it operates on semantics, and in this case the visual fidelity is what matters. What makes this even worse is that zero (I mean absolute zero) qualitative samples have been provided. Neither in the supplementary material nor in the paper to demonstrate the before / after generation quality through figures for instance. Without this signal, it is impossible to judge the validity of the proposed paper.

  5. Finally, going full circle, the motivation of this paper is absolutely unclear. Of course for something like world models, we could understand their use, but world-models are evolving to preserve explicit state that renders techniques like KV-caching unusable. All the motivation related analysis that this paper builds on top of (which is very confusing and underspecified in itself), is based on Wan 2.2 5B text-to-video generation. Since, no robust long-video generation AR-rollout candidate exists for this model, there is no clarity on exactly what problem this paper is trying to solve, and where it wishes to apply itself.

Other Minor issues:

  1. Figure 4 is confusing. Stage 1 shows computation of residuals, this confuses with the residuals of Stage 2. would be suggested to clearly delineate the two separate operations.
  2. The explanation of the Time-Adaptive allocation (B(t)) would be better prosed through an algorithm or pseudocode, instead of the current math.

Justification: Given the current state the paper is in, it's unclear motivation, and underspecified experiments, I recommend a rejection. It might be possible to work on this idea, address the weaknesses mentioned above, have better grounded motivation, in order to bring it into a good impactful state for publication. But this effort is way out of scope of the current submission, and thus I recommend reject.

Questions for authors

Would be great if you could clarify the settings for the Wan2.2 5B and the Self-Forcing 1.3B model's causal chunkwise generation. Could help understand the soundness of the experiments a bit better.

16 comments
AuthorsOpenReview ↗

Weakness 1,5 Clarification on motivation

Weakness 1,5 Clarification on motivation

Thank you for your insights and suggestions.

  1. We acknowledge that bidirectional video generation models typically do not rely on persistent KV caching. However, our work uses the Wan2.2-5B (WorldPlay AR-distilled) checkpoint, with causal attention across chunks and persistent historical KV caching, rather than the bidirectional attention used in vanilla Wan2.2-5B.
  2. KV compression already addresses a practical bottleneck in recent video-generation systems. LongLive-2.0 uses 4-bit NVFP4 KV-cache quantization, achieving approximately 3.6× cache compression.
  3. In state-of-the-art interactive world generation, LingBot-World generates video autoregressively using full-history KV caching. The full KV cache can exceed 300 GB for 1,000 generated frames.
  4. We agree that explicit world state is an important direction for world models. However, it does not eliminate the need for KV caching: Matrix-Game 4.0 maintains gameplay and physics in an explicit game engine, while its causal video renderer uses KV caching. The two are compatible.

Our motivation is that KV caching is widely used in modern autoregressive video generation and interactive world models. As generation length increases, growing KV caches increase memory usage and inference latency, motivating KV compression to improve efficiency while maintaining generation quality.

AuthorsOpenReview ↗

Weakness 2: Setting Clarafication

Clarification on experimental settings

We use the Wan2.2-5B (WorldPlay AR-distilled) checkpoint, with causal attention across chunks and persistent historical KV, instead of bidirectional attention of vanilla Wan2.2-5B. SuBitKV compresses the historical KV using offline codec calibration, with no additional generator fine-tuning.

SettingWan2.2-5BSelf-Forcing 1.3B
BaseWan2.2-TI2V-5BWan2.1-T2V-1.3B
Post-trained checkpointWorldPlay wan_distilled_model/model.ptSelf-Forcing self_forcing_dmd.pt, EMA weights
Attention / generationCausal across chunks; persistent historical KVCausal across chunks; persistent historical KV
Latent layout11 chunks × 4 frames = 4410 chunks × 3 frames = 30
RGB output173 frames, 1280 × 704117 frames, 832 × 480
AuthorsOpenReview ↗

Weakness 1: Precision reduction is not KV eviction.

SuBitKV compresses the full historical KV context, retaining a representation of every historical chunk for attention. All original compression baselines use the same full-history KV retention policy, inputs, and generation settings.

We additionally evaluated the requested eviction baselines using Wan2.2-5B (WorldPlay AR-distilled) on ten matched scene/seed pairs across three revisit gaps, under comparable memory budgets. Eviction methods retain four BF16 chunks, while SuBitKV preserves compressed representations of up to ten historical chunks.

KV memory is measured as the average final cache size over 30 runs per method, including codec data and metadata but excluding workspace and model weights.

MethodKV memory (GiB) ↓Gap 1: LPIPS ↓ / SSIM ↑ / DINO ↑Gap 2: LPIPS ↓ / SSIM ↑ / DINO ↑Gap 3: LPIPS ↓ / SSIM ↑ / DINO ↑
BF16 Full12.09– / – / 0.805– / – / 0.782– / – / 0.737
SuBitKV Q=43.920.212 / 0.7390 / 0.80540.194 / 0.770 / 0.7860.1848 / 0.7884 / 0.7479
BF16, recent-only4.830.303 / 0.607 / 0.7530.300 / 0.619 / 0.7100.308 / 0.617 / 0.659
BF16, uniform-history4.830.280 / 0.635 / 0.7740.276 / 0.657 / 0.7860.242 / 0.689 / 0.746

SuBitKV achieves better LPIPS and SSIM across all three revisit gaps while using 19% less KV memory than the eviction baselines. Its DINO scores are also comparable to or slightly higher than full-history BF16, demonstrating that preserving compressed historical information can be more effective than retaining fewer chunks at a larger memory cost.

Reviewer_wMLTOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes a training-free KV-cache compression framework for autoregressive video generation that achieves sub-bit storage while retaining an approximation of the full KV history. The method combines a persistent base constructed through RVQ, truncatable residual refinement with key-specific conditioning, and time-adaptive precision allocation based on offline calibration. Experiments on three backbones demonstrate favorable compression–quality trade-offs against existing low-bit KV-cache quantization methods.

Critical strengths
  1. The paper addresses the limitation of conventional per-element KV quantization when the available memory falls below one bit per original element.

  2. Combination of a persistent RVQ base and removable residual refinements provides a practical approach to preserving historical KV approximations under aggressive compression.

  3. Evaluation covers three backbones, multiple compression settings, prompts, and several representative quantization baselines. Component ablations, temporal diagnostics, and cache-memory accounting strengthens the empirical contribution.

Critical weaknesses
  1. Time-adaptive precision allocation treats chunk-wise distortion as separable, overlooking cross-chunk dependencies and accumulated errors during autoregressive generation.

  2. The irreversible truncation of residual refinements prevents the model from recovering precision when previously compressed historical chunks become important again.

  3. Adaptive RVQ depth determined by reconstruction MSE rather than attention sensitivity, potentially leading to suboptimal allocation of the persistent-base memory budget.

  4. Nonzero storage cost of persistent bases and reliance on a predefined generation horizon limit the framework's scalability to arbitrarily long video generation under fixed memory constraints.

Questions for authors

In Section 3.3, the allocation objective assumes that attention-weighted distortion can be expressed as a sum of individual chunk contributions. How is this assumption justified given the nonlinear interactions among historical KV states and the accumulation of quantization errors during autoregressive generation?

In Section 3.1, the adaptive RVQ depth is selected using a reconstruction-MSE threshold motivated by Gaussian rate–distortion theory. Why is this criterion appropriate for preserving attention-relevant information, given that minimizing KV reconstruction error does not necessarily minimize attention-output distortion?

The persistent base imposes a nonzero storage cost for every retained chunk, while the refinement schedule is computed offline for a predefined horizon. How does the framework guarantee the memory budget when the selected RVQ depth and corresponding base storage vary across inputs?

4 comments
AuthorsOpenReview ↗

W1 / Q1: Nonlinear Interactions and Accumulated Errors

W1 / Q1. Nonlinear interactions and accumulated errors.

Thank you for raising this important question about nonlinear interactions and error accumulation.

Let O=f(H1,…,Hn)O=f(H_1,\ldots,H_n) be the BF16 attention output for a fixed query, where Hi=(Ki,Vi)H_i=(K_i,V_i) is chunk ii's KV state. Compression introduces perturbation δi\delta_i, and Ji=∂f/∂HiJ_i=\partial f/\partial H_i is the output Jacobian evaluated at the BF16 state. For small perturbations,

ΔOall:=f(H1+δ1,…,Hn+δn)−O≈∑i=1nJiδi.\Delta O_{\mathrm{all}} := f(H_1+\delta_1,\ldots,H_n+\delta_n)-O \approx \sum_{i=1}^{n}J_i\delta_i.

The squared output error therefore satisfies

∥ΔOall∥22≈∑i=1n∥Jiδi∥22+2∑i<j⟨Jiδi,Jjδj⟩\lVert \Delta O_{\mathrm{all}} \rVert_2^2 \approx \sum_{i=1}^{n} \lVert J_i\delta_i \rVert_2^2 + 2\sum_{i<j}\langle J_i\delta_i, J_j\delta_j\rangle

The first term represents individual contributions, while the second captures cross-chunk interactions. When the cross-chunk terms are small, the individual contributions provide an additive approximation. The calibrated attention weights and distortion profiles estimate these contributions for allocation.

We test this using 2-bit refinement across ten historical chunks on identical BF16 states. Define ΔOi=f(H1,…,Hi+δi,…,Hn)−O\Delta O_i=f(H_1,\ldots,H_i+\delta_i,\ldots,H_n)-O, the measured output change when only chunk ii is compressed.

The three metrics are:

Djoint=∣ΔOall∣22∣O∣22D_{\mathrm{joint}}=\frac{\\|\Delta O_{\mathrm{all}}\\|_2^2}{\\|O\\|_2^2}

Dsum=∑i=1n∣ΔOi∣22∣O∣22D_{\mathrm{sum}}=\frac{\sum_{i=1}^{n}\\|\Delta O_i\\|_2^2}{\\|O\\|_2^2}

ϵsum=∥ΔOall−∑i=1nΔOi∥2∥ΔOall∥2\epsilon_{\mathrm{sum}} = \frac{\lVert \Delta O_{\mathrm{all}} - \sum_{i=1}^{n}\Delta O_i \rVert_2}{\lVert \Delta O_{\mathrm{all}} \rVert_2}

Results averaged over ten test scenes:

MetricResult
Joint distortion DjointD_{joint}1.402 × 10⁻³
Additive estimate DsumD_{sum}1.275 × 10⁻³
Output-change discrepancy εsumε_{sum}2.91%

The additive distortion estimate is about 9% below the measured joint distortion, while the output-change discrepancy is 2.91%, supporting the approximation in the tested setting.

We appreciate the reviewer highlighting this issue, as it motivates a more explicit analysis of the approximation.

AuthorsOpenReview ↗

W2. Irreversibility is a compression trade-off.

W2. Irreversibility is a compression trade-off.

Thank you for pointing out this important limitation.

We agree that refinement truncation is irreversible. This is a deliberate trade-off for high compression: without retaining or offloading the original high-precision states, lossy compression discards information that cannot be exactly recovered later.

SuBitKV accepts this trade-off while preserving a base representation of every historical chunk for future attention. This enables efficient long-history retention without requiring the original high-precision KV states.

AuthorsOpenReview ↗

W3 / Q2: Adaptive Depth and Attention Preservation

W3 / Q2. Adaptive depth and attention preservation.

Thank you for the insightful question about whether reconstruction MSE adequately reflects attention sensitivity.

We use reconstruction MSE rather than attention-weighted distortion for depth selection to keep the statistics lightweight and query-independent. For keys, reconstruction error bounds the squared logit perturbation by the Cauchy-Schwarz inequality:

∣qTδk∣2d≤∣q∣22∣δk∣22d.\frac{|q^T\delta k|^2}{d}\leq\frac{\\|q\\|_2^2\\|\delta k\\|_2^2}{d}.

We further validate this choice through attention-output error, comparing adaptive depth with fixed K=10 / V=8, selected on separate calibration inputs. Measurements use identical BF16 states from ten test scenes.

MeasurementAdaptiveFixed
Mean K / V depth9.887 / 7.19210 / 8
Joint K/V attention-output NMSE ↓0.062930.07323

Adaptive depth uses fewer stages and reduces joint attention-output error by 14.1%, demonstrating the effectiveness of this lightweight criterion in the tested setting.

We thank the reviewer for encouraging this additional validation.

AuthorsOpenReview ↗

W4 / Q3. Full-history compression and bounded base storage.

W4 / Q3. Full-history compression and bounded base storage.

Thank you for raising the important question of long-term memory growth and bounded storage.

SuBitKV compresses the complete visible KV history, retaining a representation of every historical chunk for attention. All original compression baselines share the same full-history KV retention policy, inputs, and generation settings.

With nonzero storage per chunk, retaining an infinite history under fixed memory is impossible without discarding history. KV compression and eviction are orthogonal and can be combined: compression reduces storage per retained chunk, while eviction controls which chunks remain.

Base depth is empirically stable across the ten test scenes: the mean combined K+V RVQ depth is 17.079 stages [99.5% CI of the mean: 16.717, 17.423] (K=9.887, V=7.192).

We also impose a hard maximum of 16 RVQ stages for each K or V base, giving an input-independent bound on per-chunk base storage and, for a specified history horizon, total base storage.

We appreciate the reviewer highlighting this distinction between compression efficiency and long-term memory management.

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

SuBitKV compresses video-generation KV caches using an adaptive residual-vector-quantized base and progressively truncatable residual bits. An offline schedule allocates precision across cache ages under a finite-horizon memory budget. Experiments on three backbones evaluate reconstruction, video quality and storage, with up to 30.09× compression on Wan2.2.

Critical strengths
  1. The persistent base and removable refinement bits fit the problem well: old states can lose precision without disappearing from attention. The matched-storage ablation is particularly convincing. At roughly 15× compression, adding the base improves LPIPS from 0.344 to 0.276, isolating a real benefit of the representation.
  2. The measured trade-off is strong. On Wan2.2, Q = 0.2 reaches 21.42× compression with LPIPS 0.327, compared with 5.33× and 0.388 for INT2 KIVI. The reported Wan runtime overhead is small enough to make the method practically interesting.
Critical weaknesses
  1. The Wan2.2-5B evaluation protocol is underspecified. The official model uses bidirectional attention and jointly denoises the video, whereas SuBitKV assumes persistent KV caches across autoregressive chunks. What checkpoint and inference modifications enable this evaluation?
  2. Retaining every historical chunk is not shown to preserve useful long-term information. The comparisons fix the retention policy, and the main Wan evaluation covers only 173 frames. These experiments establish compression fidelity, but do not test whether storing all old chunks at very low precision is better than keeping a smaller, more accurate history under the same memory budget. Because the persistent base still grows with history length, the practical justification depends on this trade-off. A test requiring information to survive a long gap is missing.
Questions for authors
  1. Under the same peak memory budget, does preserving all compressed chunks outperform an anchor-plus-recent-window baseline with higher-precision KV on object reappearance or camera revisit sequences? How does the gap change as the revisit interval grows? This directly tests whether the retained old cache remains useful.
3 comments
AuthorsOpenReview ↗

Clarification on Experimental Settings

Clarification on Experimental Settings

Thank you for the constructive feedback and for the opportunity to clarify our experimental setup.

We use the Wan2.2-5B (WorldPlay AR-distilled) checkpoint, with causal attention across chunks and persistent historical KV caching, rather than the bidirectional attention used in vanilla Wan2.2-5B. SuBitKV compresses the historical KV using offline codec calibration, with no additional generator fine-tuning.

SettingWan2.2-5B
BaseWan2.2-TI2V-5B
Post-trained checkpointWorldPlay wan_distilled_model/model.pt
Attention / generationCausal across chunks; persistent historical KV
Latent layout11 chunks × 4 frames = 44
RGB output173 frames, 1280 × 704

We appreciate the reviewer's suggestion and will clarify these implementation details in the revised paper.

AuthorsOpenReview ↗

More Compressed History vs. Fewer High-Precision Chunks

More Compressed History vs. Fewer High-Precision Chunks

Thank you for this valuable suggestion. We conducted additional experiments to directly compare full-history KV compression with retaining fewer high-precision chunks.

SuBitKV compresses the complete visible KV history, retaining a representation of every historical chunk for attention. All original compression baselines share the same full-history KV retention policy, inputs, and generation settings.

We additionally evaluated the requested eviction baselines using Wan2.2-5B (WorldPlay AR-distilled) on ten matched scene/seed pairs across three revisit gaps, under comparable memory budgets. Eviction methods retain four BF16 chunks, while SuBitKV preserves compressed representations of up to ten historical chunks.

KV memory is measured as the average final cache size over 30 evaluation runs per method, including codec data and metadata but excluding workspace and model weights.

MethodKV memory (GiB) ↓Gap 1: LPIPS ↓ / SSIM ↑ / DINO ↑Gap 2: LPIPS ↓ / SSIM ↑ / DINO ↑Gap 3: LPIPS ↓ / SSIM ↑ / DINO ↑
BF16 Full (reference)12.0850.0000 / 1.0000 / 0.80490.0000 / 1.0000 / 0.78180.0000 / 1.0000 / 0.7368
SuBitKV Q=4, all history3.9170.2124 / 0.7390 / 0.80540.1936 / 0.7695 / 0.78580.1848 / 0.7884 / 0.7479
BF16, recent-only4.8340.3029 / 0.6074 / 0.75250.2995 / 0.6192 / 0.71010.3075 / 0.6170 / 0.6592
BF16, uniform-history4.8340.2800 / 0.6354 / 0.77370.2756 / 0.6570 / 0.78610.2423 / 0.6893 / 0.7463

SuBitKV achieves better LPIPS and SSIM across all three revisit gaps while using approximately 19% less KV memory than the eviction baselines, with DINO scores comparable to full-history BF16.

These results demonstrate that retaining compressed representations of more historical chunks can better preserve revisit consistency than retaining fewer high-precision chunks, even with a smaller memory footprint.

We thank the reviewer for suggesting this comparison, which helps further validate the benefits of full-history KV compression.

AuthorsOpenReview ↗

Additional Results Under Tighter KV Memory Budgets

Thank you for the valuable suggestion. To further examine the trade-off between preserving compressed history and retaining fewer high-precision chunks, we extended our experiments to more aggressive memory budgets.

We use the same experimental settings as in our previous response: the Wan2.2-5B (WorldPlay AR-distilled) checkpoint, ten matched scene/seed pairs, 173-frame videos, and revisit gaps of 1/2/3 chunks. PSNR, SSIM, and LPIPS are evaluated over the ten complete chunks following the shared initial chunk, using full-history BF16 as the reference. PSNR averages per-frame dB values, with zero-error frames capped at 120 dB; BF16 self-comparison is reported as ∞. DINO measures within-video revisit similarity.

MethodKV Memory (GiB) ↓Gap 1: PSNR ↑ / SSIM ↑ / LPIPS ↓ / DINO ↑Gap 2: PSNR ↑ / SSIM ↑ / LPIPS ↓ / DINO ↑Gap 3: PSNR ↑ / SSIM ↑ / LPIPS ↓ / DINO ↑
BF16 Full (reference)12.085∞ / 1.0000 / 0.0000 / 0.8049∞ / 1.0000 / 0.0000 / 0.7818∞ / 1.0000 / 0.0000 / 0.7368
SuBitKV Q=21.95221.81 / 0.6688 / 0.2661 / 0.804422.15 / 0.6808 / 0.2573 / 0.788422.46 / 0.6976 / 0.2456 / 0.7443
SuBitKV Q=11.19718.74 / 0.5913 / 0.3226 / 0.802718.82 / 0.5935 / 0.3297 / 0.777119.05 / 0.6014 / 0.3242 / 0.7181
BF16 recent-22.41718.66 / 0.4390 / 0.4531 / 0.671618.47 / 0.4354 / 0.4701 / 0.611118.49 / 0.4336 / 0.4733 / 0.5945
BF16 uniform-22.41719.99 / 0.4721 / 0.4160 / 0.786120.01 / 0.4752 / 0.4158 / 0.744620.13 / 0.4808 / 0.4090 / 0.7399
BF16 recent-11.20913.57 / 0.3523 / 0.5357 / 0.716413.39 / 0.3438 / 0.5471 / 0.657913.36 / 0.3424 / 0.5492 / 0.6154

SuBitKV Q=2 outperforms both two-chunk BF16 baselines across all four metrics and all three revisit gaps, while using approximately 19% less KV memory. Compared with uniform two-chunk BF16 retention, SuBitKV Q=2 reduces LPIPS by 36.0%, 38.1%, and 40.0% at revisit gaps of 1, 2, and 3 chunks, respectively.

Furthermore, SuBitKV Q=1 retains compressed representations of all historical chunks using only 1.197 GiB, slightly less than the 1.209 GiB required for a single BF16 chunk, while achieving better PSNR, SSIM, and LPIPS across all three revisit gaps.

These additional results complement our previous four-chunk comparison and further demonstrate the benefits of retaining compressed historical KV under increasingly constrained memory budgets. The improvements persist across longer revisit intervals, supporting the usefulness of retained historical information.

We thank the reviewer again for the constructive feedback and would be happy to provide further clarification if needed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Follow-up on Rebuttal and Updated Materials

Dear Reviewer,

We would like to kindly follow up on our rebuttal. We have also uploaded a revised PDF and supplementary video materials to further address your concerns.

We would greatly appreciate any additional feedback you may have and would be happy to provide further clarification if needed.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.