Pith. sign in

REVIEW 5 major objections 5 minor 22 references

For visual autoregressive image models, the entropy of each token's predicted distribution can be read online to decide which tokens to skip, yielding up to a 2.89x speedup with negligible quality loss.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 05:43 UTC pith:UYZMEVOE

load-bearing objection NOVA is a genuinely adaptive token-pruning method for VAR with solid ablations, but the theoretical justification is overstated and the evaluations need error bars. the 5 major comments →

arxiv 2602.01345 v2 pith:UYZMEVOE submitted 2026-02-01 cs.CV

Adaptive Visual Autoregressive Acceleration via Dual-Linkage Entropy Analysis

Yu Zhang , Jingyi Liu , Feng Liu , Duoqian Miao , Qi Zhang , Kexue Fu , Changwei Wang , Longbing Cao This is my paper
classification cs.CV
keywords visual autoregressive modelstoken reductionentropy-guided pruninginference accelerationtraining-free accelerationtext-to-image generationresidual cache reuselatency reduction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that the predictive entropy of each token during visual autoregressive (VAR) generation reveals which tokens are safe to skip, and that this signal can be read online to accelerate inference without retraining. It introduces NOVA, which detects when scale-level entropy growth slows (an inflection point) then prunes low-entropy tokens with per-scale and per-layer ratios, reusing cached residuals to fill in the dropped tokens. On standard benchmarks NOVA reports a 2.89x speedup for Infinity-2B with essentially unchanged GenEval accuracy, and on ImageReward it slightly beats the unpruned Infinity-8B in preference score while cutting latency roughly in half. A sympathetic reader would care because if correct, it means a cheap, content-adaptive scheduler can unlock much of the theoretical redundancy in VAR models without any fine-tuning.

Core claim

NOVA is a training-free token-reduction framework that treats a token's predictive entropy H_{t,i} (the entropy of its predicted class distribution given previous scales) as the criterion for pruning. It computes the mean scale-level entropy, detects the first scale where the smoothed growth rate drops below a fraction α of the early baseline, and uses that as the acceleration activation scale. It then assigns each scale a base reduction ratio that grows with scale index, adjusted by the current entropy growth rate, and refines that ratio per layer using the relative deviation of the layer's mean entropy from the mean of preceding layers. Pruned tokens are reconstructed by interpolating the

What carries the argument

The load-bearing object is the token-level predictive entropy H_{t,i} = -Σ_v p_{t,i}(v) log p_{t,i}(v), measured at every scale and layer during generation. It serves two roles: as the online signal for detecting the entropy-growth inflection point (via smoothed discrete growth rate g_t and baseline η) that triggers acceleration, and as the criterion for which tokens to prune, with per-scale and per-layer ratios set by sigmoid and tanh functions of scale index and entropy growth. Residual cache reuse (interpolating Cache_{t,j} = R_output_{t,j} - R_input_{t,j} from scale t to t+1) reconstructs pruned positions, preserving structure without recomputation.

Load-bearing premise

The claim rests on the premise that a token's predictive entropy at the current scale reliably identifies tokens whose removal costs little future generation quality — even though the theory only bounds what high-entropy tokens could carry, not what low-entropy tokens actually contribute.

What would settle it

A concrete test: run NOVA on a diverse set of prompts where the object has fine, load-bearing structures (e.g., text, faces, tree branches) and measure GenEval/HPS against the original model; if there exist categories where the low-entropy pruned tokens are exactly those carrying structural anchors and quality drops sharply, the entropy criterion fails. Alternatively, construct a synthetic VAR process where the low-entropy tokens are decisive for later scales, and show the speedup/quality trade-off inverts.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If correct, any VAR-style text-to-image model can be accelerated by a training-free entropy readout, with bigger gains on larger models where latency dominates.
  • The adaptive activation scale removes the need for hand-tuned stage boundaries; the schedule adapts per image category, suggesting content-dependent acceleration.
  • The ImageReward result (1.035 vs 1.030) implies token pruning need not degrade human preference, which challenges the assumption that acceleration always trades off quality.
  • Residual cache reuse provides a template for combining token sparsity with cache-based reconstruction in other dense prediction tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The entropy signal could be re-used for other decisions, such as adaptive sampling steps in diffusion or deciding when to stop refinement, since it measures where uncertainty remains.
  • The hand-set hyperparameters (α=0.5, test=5, τ=0.8, λ=0.1) are a search space; one could auto-tune them per model, or test the robustness across a wider set of prompts and random seeds than reported.
  • Because the theoretical bound is one-sided, the biggest risk is that some low-entropy tokens are structurally important anchors; identifying such failure cases (e.g., object boundaries, text glyphs) would sharpen the method.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes NOVA, a training-free token-reduction acceleration framework for visual autoregressive (VAR) text-to-image models. NOVA uses per-token predictive entropy to identify low-information tokens, detects an instance-dependent 'inflection point' in scale-level entropy growth to decide when acceleration starts, computes per-scale and per-layer reduction ratios via scale- and layer-linkage functions, and reuses residual caches for pruned tokens. The authors report a 2.89× speedup on Infinity-2B with near-lossless GenEval performance and a 1.33× speedup on Infinity-8B with a slightly higher ImageReward score than the unpruned baseline. The paper includes ablations on the token-selection metric and on the linkage components.

Significance. Training-free acceleration for VAR is practically relevant, and the reported speedups, if robust, would be a useful contribution: NOVA adds adaptivity to a class of methods that typically rely on fixed schedules, and entropy-based token selection is a principled alternative to attention- or frequency-based heuristics. The paper provides quantitative comparisons on several backbones and benchmarks, as well as ablations for the main design choices. However, the theoretical grounding as stated does not establish pruning safety, and several algorithmic definitions are incomplete, so the significance is conditional on revision.

major comments (5)
  1. [Eq. (12), §3.3] H̄_{t,k}, the per-layer mean entropy, is used in Eqs. (12)–(13) but never defined. Eq. (7) defines H̄_t only for a whole token map R_t, from the token predictive entropies p_{t,i}. Since the layer-linkage adjustment operates per Transformer layer, the paper must specify which intermediate output produces the per-layer entropy or how H̄_{t,j} is computed. Without this, the layer-adaptive ratios are not implementable or reproducible.
  2. [§3.2, Eqs. (2)–(3)] The displayed inequalities bound the amount of information a low-entropy token could carry, but pruning does not merely ignore the token's discrete value; it deletes the hidden representation from later layers and substitutes a cached residual (Eqs. 14–16). The relevant quantity for safety is the divergence between pruned and unpruned future-scale predictive distributions, which is not bounded by H(X|Z). The statement in §3.2 that pruning 'has limited impact on potential information gain' is therefore not a theorem. Please add a bound on the actual distribution shift or explicitly reclassify the entropy criterion as an empirical heuristic supported by the ablations.
  3. [Table 4, §4.3] The central near-lossless claim rests on a single ablation (Infinity-2B, GenEval) with point estimates and no variance or significance testing. The headline differences (GenEval 0.72 vs 0.73; ImageReward 1.035 vs 1.030) are within typical measurement noise. Please report multiple seeds or confidence intervals for at least the key configurations, and ideally an additional backbone/benchmark, to substantiate the safety of the entropy criterion.
  4. [Eq. (16), §3.3] The residual cache reuse equation is written as Routput_{t,j} = Cache_{t→t+1,j} + Rinput_{t,j}, but the left-hand side should refer to the next scale (t+1), otherwise the equation is circular. Please correct the indexing. Relatedly, Eq. (13) can yield negative ratios when Δ_{t,j} > 1; specify clipping or other bounds on the reduction ratio.
  5. [§4.1, hyperparameters] Four hand-set hyperparameters (test=5, α=0.5, τ=0.8, λ=0.1) are used without sensitivity analysis. Because the paper emphasizes adaptivity rather than tuning, please show how results vary with these settings (e.g., a small grid) to demonstrate that the method is not brittle to their choice.
minor comments (5)
  1. [Eq. (10), §3.3] The threshold definition uses t_est, which is never introduced; presumably this should be test, the baseline window parameter. Please define it.
  2. [Abstract and §4.2] The abstract says '0.01% performance loss' but Table 1 shows GenEval Overall 0.72 vs 0.73, which is about 1.4% relative. Please clarify whether '0.01' refers to an absolute point difference or a relative percentage.
  3. [Eqs. (5)–(6), §3.3] The factorized likelihood assumption within a scale is central to the additive entropy decomposition. State explicitly that this is the model's parallel-prediction assumption and discuss any mismatch with actual VAR implementations that use causal masking within a scale.
  4. [Figure 4(b), §3.2] The caption says 600 prompts from GenEval were 'rewritten' but no details of the rewriting are given. Please provide prompt-processing details for reproducibility.
  5. [Notation] The manuscript alternates between 'NOVA' and 'NOV A'; standardize the name. Also, the reference list has an entry 'Li, Y., Wang, H., et al.' in §2 with missing year/venue details.

Circularity Check

0 steps flagged

No circularity: NOVA's speed/quality claims are validated on external benchmarks, and the entropy-based selection rule is a heuristic with no input-output equivalence.

full rationale

The central claim (e.g., 2.89x speedup on Infinity-2B with ~0.01 GenEval loss) is not derived from entropy by construction. Token entropy H_{t,i} (Eq. 4) is computed from the model's predictive distribution; the reduction ratios (Eqs. 11 and 13) use entropy and hand-set hyperparameters (α=0.5, τ=0.8, λ=0.1), but the resulting quality is then measured against external benchmarks (GenEval, DPG-Bench, MJHQ-30K, ImageReward) in Tables 1-3. Thus the reported outcomes are independent of the method's internal definitions. The theoretical support (Eq. 2) is a one-sided bound that does not by itself prove pruning safety; however, a weak or invalid inference is a correctness concern, not circularity. The paper's self-citations (e.g., Zhang et al. 2025b in Related Work) are background-only and do not carry any load-bearing argument. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work. The ablation in Table 4 empirically compares entropy against attention- and MSE-based selection on the same benchmark, so the choice of criterion is tested rather than assumed. Consequently, no specific equation reduces to its own input, and no circular step can be quoted with a concrete reduction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The ledger is light: NOVA introduces no new entities and relies on standard information-theoretic inequalities, but its operational decisions ride on four hand-set constants (α, test, τ, λ) and on three empirical/domain assumptions (the entropy two-stage dynamic, the low-entropy-safety leap, and the residual-cache proxy) that are validated only by internal ablations on a single benchmark set.

free parameters (4)
  • α (entropy growth threshold) = 0.5
    Eq (10): acceleration activates when the smoothed entropy growth rate g̃t drops below α·η. The threshold directly sets the speed/quality tradeoff and is stated in §4.1 without sensitivity analysis; no ablation varies it.
  • test (early baseline window) = 5
    Eq (9): the first test scales are never pruned and define the baseline growth η. test=5 hard-codes a minimum latency before any acceleration, partially undercutting the claim of solving 'limited acceleration scope'.
  • τ (sigmoid temperature) = 0.8
    Eq (11): scale-linkage ratio Ratiot = σ((t−(t⋆+1))/τ) − λ·tanh(gt). Sets how quickly the base pruning ratio ramps up; no sensitivity analysis.
  • λ (tanh scaling) = 0.1
    Eq (11): scales the entropy-growth-rate modulation of the pruning ratio; hand-set without sensitivity analysis.
axioms (4)
  • domain assumption Factorized conditional likelihood within each scale: p(Rt|R<t) = ∏ p(rt,i|R<t) (Eq 5)
    Inherited from VAR's parallel next-scale prediction design; it makes Eq (6)'s additive entropy decomposition exact, but it is the model's independence assumption, not a fact about the true joint distribution of the image.
  • domain assumption Two-stage entropy dynamic: scale-mean entropy grows rapidly then plateaus for every instance
    Empirical claim from Figure 4(b) over rewritten GenEval prompts; the whole activation rule (Eqs 8–10) presumes this shape. Instances with flat or oscillating entropy growth would activate pruning at the wrong scale.
  • ad hoc to paper Low-entropy tokens contribute negligible future information gain (safety of pruning)
    The derivation in §3.2 gives only I(X;Y|Z) ≤ H(X|Z) (Eq 2), a one-sided bound; the leap that pruning the lowest-entropy tokens has 'limited impact on potential information gain' does not follow from the bound. Support is the empirical ablation in Table 4, not the displayed inequalities.
  • domain assumption Residual cache from the prior scale interpolates to a faithful substitute for the current scale's layer output on pruned tokens
    Eqs (14)–(16): the approximation error of reusing the previous scale's residual is never analyzed; its validity is assumed to hold only for pruned (low-entropy) tokens.

pith-pipeline@v1.3.0-alltime-deepseek · 12370 in / 24585 out tokens · 239707 ms · 2026-08-03T05:43:16.540818+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Adaptive Visual Autoregressive Acceleration via Dual-Linkage Entropy Analysis." pith.science (2026). https://pith.science/paper/UYZMEVOE

@misc{pith2026260201345,
  author       = {Pith},
  title        = {Pith review of: Adaptive Visual Autoregressive Acceleration via Dual-Linkage Entropy Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UYZMEVOE}},
  note         = {Machine review of arXiv:2602.01345}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Visual AutoRegressive modeling (VAR) suffers from substantial computational cost due to the massive token count involved. Failing to account for the continuous evolution of modeling dynamics, existing VAR token reduction methods face three key limitations: heuristic stage partition, non-adaptive schedules, and limited acceleration scope, thereby leaving significant acceleration potential untapped. Since entropy variation intrinsically reflects the transition of predictive uncertainty, it offers a principled measure to capture modeling dynamics evolution. Therefore, we propose NOVA, a training-free token reduction acceleration framework for VAR models via entropy analysis. NOVA adaptively determines the acceleration activation scale during inference by online identifying the inflection point of scale entropy growth. Through scale-linkage and layer-linkage ratio adjustment, NOVA dynamically computes distinct token reduction ratios for each scale and layer, pruning low-entropy tokens while reusing the cache derived from the residuals at the prior scale to accelerate inference and maintain generation quality. Extensive experiments and analyses validate NOVA as a simple yet effective training-free acceleration framework.

Figures

Figures reproduced from arXiv: 2602.01345 by Changwei Wang, Duoqian Miao, Feng Liu, Jingyi Liu, Kexue Fu, Longbing Cao, Qi Zhang, Yu Zhang.

Figure 1
Figure 1. Figure 1: NOVA generates a 2K image in 3.2 seconds on a single NVIDIA RTX 3090 GPU without running out of GPU memory. Abstract Visual AutoRegressive modeling (VAR) suffers from substantial computational cost due to the massive token count involved. Failing to account for the continuous evolution of modeling dynam￾ics, existing VAR token reduction methods face three key limitations: heuristic stage partition, non-ada… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Peak GPU memory consumption across different scales during 1024×1024 image generation in VAR. As the scale increases, the token sequence length grows rapidly, leading to quadratic growth in attention computation complexity. (b) Theo￾retical GFLOPs required by next-token and next-scale prediction models to generate images of various resolutions under uniform settings. Compared to the raster-scan next-to… view at source ↗
Figure 4
Figure 4. Figure 4: (a) Frequency spectrum of various instances at different scales during VAR generation. (b) Various entropy values across different scales during the inference process of Infinity-2B: We selected 600 prompts from GenEval and rewrote them. For each rewritten prompt, we generated four images using different random seeds. (c) Entropy heatmaps of token maps across different Transformer layers. Cross-layer spati… view at source ↗
Figure 5
Figure 5. Figure 5: NOVA framework at both scale-level and layer-level. where pt,i(·) is the predicted class distribution for token rt,i. Since VAR performs parallel prediction for Nt tokens within the t-th scale, we adopt a factorized conditional likelihood within the token map Rt: p(Rt | R<t) = Y Nt i=1 p(rt,i | R<t). (5) Therefore, the joint predictive entropy of the token map Rt decomposes additively into token predictive… view at source ↗
Figure 6
Figure 6. Figure 6: Visual comparison between NOVA and other training￾free comparable methods on Infinity-2B. Red and yellow boxes highlight the finer details in visualization. under the adaptive acceleration activation and dual-linkage acceleration, NOVA achieves near-lossless visual quality mitigates semantic loss, structural distortion, and detail col￾lapse. Notably, in the yellow-box region, NOVA generates clearer and mor… view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparisons between NOVA and backbones, including Infinity-2B, Infinity-8B, and HART-0.7B. we conduct an ablation study comparing it with Attention￾based (Attn) and MSE-based (MSE) alternatives on Infinity￾2B. All variants adopt the same dual-linkage configuration for fair comparison. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Distribution of acceleration activation scale ratios of NOVA across different categories of the MJHQ-30K benchmark. set depends on content: fine-grained textures necessitate later pruning, whereas simple global structures allow earlier activation. The results suggest that NOVA can effectively capture context-specific generation characteristics and adap￾tively identify appropriate acceleration scales, rathe… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 12 linked inside Pith

  1. [1]

    Accelerated sampling from masked diffusion models via entropy bounded unmasking.arXiv preprint arXiv:2505.24857,

    Ben-Hamu, H., Gat, I., Severo, D., Nolte, N., and Kar- rer, B. Accelerated sampling from masked diffusion models via entropy bounded unmasking.arXiv preprint arXiv:2505.24857,

    Pith/arXiv arXiv

  2. [5]

    Ella: Equip diffusion models with llm for enhanced semantic alignment.arXiv preprint arXiv:2403.05135,

    Hu, X., Wang, R., Fang, Y ., Fu, B., Cheng, P., and Yu, G. Ella: Equip diffusion models with llm for enhanced semantic alignment.arXiv preprint arXiv:2403.05135,

    Pith/arXiv arXiv

  3. [6]

    Spectralar: Spectral autoregressive visual generation

    Huang, Y ., Chen, W., Zheng, W., Duan, Y ., Zhou, J., and Lu, J. Spectralar: Spectral autoregressive visual generation. arXiv preprint arXiv:2506.10962,

    Pith/arXiv arXiv

  4. [7]

    Flexvar: Flexible visual au- toregressive modeling without residual prediction.arXiv preprint arXiv:2502.20313,

    Jiao, S., Zhang, G., Qian, Y ., Huang, J., Zhao, Y ., Shi, H., Ma, L., Wei, Y ., and Jie, Z. Flexvar: Flexible visual au- toregressive modeling without residual prediction.arXiv preprint arXiv:2502.20313,

    arXiv

  5. [8]

    G., and Patel, V

    Kumar, A., Nair, N. G., and Patel, V . M. Scale- wise var is secretly discrete diffusion.arXiv preprint arXiv:2509.22636,

    arXiv

  6. [9]

    Efficient training for human video generation with entropy-guided prioritized progressive learning.arXiv preprint arXiv:2511.21136, 2025a

    Li, C., Zhang, J., Liu, S., Lin, S., Shi, Z., Li, Z., and Chang, X. Efficient training for human video generation with entropy-guided prioritized progressive learning.arXiv preprint arXiv:2511.21136, 2025a. Li, D., Kamko, A., Akhgari, E., Sabet, A., Xu, L., and Doshi, S. Playground v2. 5: Three insights towards enhancing aesthetic quality in text-to-image...

    arXiv

  7. [10]

    Improving the trainability of deep neural networks through layerwise batch-entropy regularization

    Peer, D., Keulen, B., Stabinger, S., Piater, J., and Rodr´ıguez- S´anchez, A. Improving the trainability of deep neural networks through layerwise batch-entropy regularization. arXiv preprint arXiv:2208.01134,

    Pith/arXiv arXiv

  8. [12]

    Visual autoregressive modeling for image super- resolution.arXiv preprint arXiv:2501.18993,

    Qu, Y ., Yuan, K., Hao, J., Zhao, K., Xie, Q., Sun, M., and Zhou, C. Visual autoregressive modeling for image super- resolution.arXiv preprint arXiv:2501.18993,

    Pith/arXiv arXiv

  9. [13]

    Rajagopalan, S., Narayan, K., and Patel, V . M. Restore- var: Visual autoregressive generation for all-in-one image restoration.arXiv preprint arXiv:2505.18047,

    arXiv

  10. [14]

    M-var: Decoupled scale-wise autoregressive model- ing for high-quality image generation.arXiv preprint arXiv:2411.10433,

    Ren, S., Yu, Y ., Ruiz, N., Wang, F., Yuille, A., and Xie, C. M-var: Decoupled scale-wise autoregressive model- ing for high-quality image generation.arXiv preprint arXiv:2411.10433,

    Pith/arXiv arXiv

  11. [15]

    Efficient diffusion models: A survey.arXiv preprint arXiv:2502.06805,

    Shen, H., Zhang, J., Xiong, B., Hu, R., Chen, S., Wan, Z., Wang, X., Zhang, Y ., Gong, Z., Bao, G., et al. Efficient diffusion models: A survey.arXiv preprint arXiv:2502.06805,

    Pith/arXiv arXiv

  12. [16]

    Hart: Efficient visual generation with hybrid autoregressive transformer

    Tang, H., Wu, Y ., Yang, S., Xie, E., Chen, J., Chen, J., Zhang, Z., Cai, H., Lu, Y ., and Han, S. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812,

    Pith/arXiv arXiv

  13. [17]

    Dynamic mixture-of- experts for visual autoregressive model.arXiv preprint arXiv:2510.08629,

    Vincenti, J., Jazbec, M., and Xia, G. Dynamic mixture-of- experts for visual autoregressive model.arXiv preprint arXiv:2510.08629,

    arXiv

  14. [18]

    Training-free text-guided image edit- ing with visual autoregressive model.arXiv preprint arXiv:2503.23897,

    Wang, Y ., Guo, L., Li, Z., Huang, J., Wang, P., Wen, B., and Wang, J. Training-free text-guided image edit- ing with visual autoregressive model.arXiv preprint arXiv:2503.23897,

    Pith/arXiv arXiv

  15. [19]

    Wu, J., Chen, X., Shi, B., Li, S., and Xu, K

    URL https://arxiv.org/ abs/2510.23028. Wu, J., Chen, X., Shi, B., Li, S., and Xu, K. Sega: Structural entropy guided anchor view for graph contrastive learning. InInternational Conference on Machine Learning, pp. 37293–37312. PMLR, 2023a. Wu, X., Hao, Y ., Sun, K., Chen, Y ., Zhu, F., Zhao, R., and Li, H. Human preference score v2: A solid benchmark for e...

    arXiv

  16. [20]

    Litevar: Compressing visual autoregres- sive modelling with efficient attention and quantization

    Xie, R., Zhao, T., Yuan, Z., Wan, R., Gao, W., Zhu, Z., Ning, X., and Wang, Y . Litevar: Compressing visual autoregres- sive modelling with efficient attention and quantization. arXiv preprint arXiv:2411.17178,

    Pith/arXiv arXiv

  17. [21]

    Actvar: Activating mixtures of weights and tokens for efficient visual autoregressive generation.arXiv preprint arXiv:2511.12893, 2025a

    Zhang, K., Yang, R., Zhang, Y ., You, S., and Huang, T. Actvar: Activating mixtures of weights and tokens for efficient visual autoregressive generation.arXiv preprint arXiv:2511.12893, 2025a. 10 Adaptive Visual Autoregressive Acceleration via Dual-Linkage Entropy Analysis Zhang, Q., Dai, X., Yang, N., An, X., Feng, Z., and Ren, X. Var-clip: Text-to-image...

    arXiv

  18. [22]

    Markovian scale prediction: A new era of visual autoregressive generation.arXiv preprint arXiv:2511.23334, 2025b

    Zhang, Y ., Liu, J., Shi, Y ., Zhang, Q., Miao, D., Wang, C., and Cao, L. Markovian scale prediction: A new era of visual autoregressive generation.arXiv preprint arXiv:2511.23334, 2025b. 11

    arXiv

  19. [2022]

    Head-aware kv cache compression for ef- ficient visual autoregressive modeling.arXiv preprint arXiv:2504.09261,

    Qin, Z., Lv, Y ., Lin, M., Guo, H., Zhang, Z., Zou, D., and Lin, W. Head-aware kv cache compression for ef- ficient visual autoregressive modeling.arXiv preprint arXiv:2504.09261,

    arXiv

  20. [2023]

    Fastvar: Linear visual autoregres- sive modeling via cached token pruning.arXiv preprint arXiv:2503.23367, 2025a

    Guo, H., Li, Y ., Zhang, T., Wang, J., Dai, T., Xia, S.- T., and Benini, L. Fastvar: Linear visual autoregres- sive modeling via cached token pruning.arXiv preprint arXiv:2503.23367, 2025a. Guo, L., Xue, Z., Xu, Z., and Xu, H. Demospeedup: Accel- erating visuomotor policies via entropy-guided demon- stration acceleration.arXiv preprint arXiv:2506.05064, 2...

    Pith/arXiv arXiv

  21. [2024]

    Tts-var: A test-time scaling framework for visual auto-regressive generation.arXiv preprint arXiv:2507.18537, 2025a

    Chen, Z., Chu, R., Chen, Y ., Zhang, S., Wei, Y ., Zhang, Y ., and Liu, X. Tts-var: A test-time scaling framework for visual auto-regressive generation.arXiv preprint arXiv:2507.18537, 2025a. Chen, Z., Fan, J., Yu, Z., Zhuang, B., and Tan, M. Frequency-aware autoregressive modeling for efficient high-resolution image synthesis. InProceedings of the IEEE/C...

    arXiv

  22. [2025]

    Neighboring autoregressive modeling for effi- cient visual generation.arXiv preprint arXiv:2503.10696,

    He, Y ., He, Y ., He, S., Chen, F., Zhou, H., Zhang, K., and Zhuang, B. Neighboring autoregressive modeling for effi- cient visual generation.arXiv preprint arXiv:2503.10696,

    Pith/arXiv arXiv