- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3(+2)
Read review
Paper summary
The major contribution of this work is to comprehensively address the practical deployment issue of model weight compression for arbitrary bitrates. It clearly outlines the challenges and properly handle them by stacking row normalization, the construction of probability model with heuristic estimation method. GPU parallelization is also properly addressed in their decoding design.
Critical strengths
Mentioned in the summary.
Critical weaknesses
Overall, this work is built by the stacking of several existing methods or concepts with its unique design, being deployed together in an organized way, and being validated on modern models. This makes the evaluation section the section we heavily rely on, since the major contribution is the engineering and the result matters. However, the current state of evaluation is too thin and cannot properly make the argument of such a work solid. I list all the weaknesses in the Questions for Authors section.
Questions for authors
- My major concern to address is how does each component in this work stack with each other, from the perspective of the performance portfolio (including the stored rate, relative weight error, task fidelity, model conversion time, and inference time), and what is the takeway of each component. This is important since if this is properly address, we can have clear insight on why each component is designed in their ways.
- There is a lack of ablation study on each design block. How much does it help if the RDO is enabled or disabled? For the stored size estimation, if we don’t estimate, but directly use the ground-truth? This will consume much more computation and therefore is not practically deployable, but this ablation can help us realize the trade-off between the computation and the estimation.
- The evaluation data is too thin, and not documented in a reproducible manner. I only see thin clue in Appendix about what constitutes the data (what prompts are used) for table 1 and 2. In line 680, it is mentioned that “Z-Image-Turbo generations use eight denoising steps, two prompts, and seeds 42, 7, and 123.” So, which two prompts are used? Most importantly, two are just too tiny. Two prompts is not enough to be considered for the full figure 1 and table 1’s performance. This makes the whole analysis in danger. Similarly, the prompt used for LLM also looks extremely thin. “We evaluate Qwen3.8-27B perplexity on WikiText-2 raw test text”. How long is this text dataset? What is the coverage of this dataset?
- Only one candidate model for vision and language model each is evaluated. How does this method scale with model size. For example, Qwen3.8-27B is tested for LLM, then how about other model sizes in Qwen series? How about other families of open-weight LLM? Without evaluation, it is hard to tell whether this method actually outperform SOTA for general deployment. This is important since the current result doesn’t stand out by a meaningful margin, compared to methods like HyperQuant, SDNQ, and HRTN, especially for MiniMax-H3.
- The design of equation 10 is more like heuristics. Do you have theoretical backup for this design? Also, how accurate is the estimation of equation 12, compared to ground-truth?
- Table 1 and 2 are too noisy and mixed. All the results with different bpp are mixed together, and therefore hard to comprehend what is the Pareto frontier this method occupy. Please put some effort in this part and structure the visualization and demonstration.
4 comments
3. Replies to “The evaluation data is too thin” and limited model coverage (Q3 & Q4)
Following your feedback, we redesigned the main evaluation and expanded its coverage to Z-Image-Turbo and Qwen-Image-2.1-Turbo for image generation, MiniMax-H3 for video and audio generation, and Qwen3.8-27B, Muse-Glimmer-30B, and Qwen3.5-9B for language modeling. These include the different model sizes and families you requested. We evaluate them on widely used benchmarks: GenEval and DPG-Bench for images, VBench for video, and WikiText-2, MMLU, GSM8K, HellaSwag, and ARC-Challenge for language. PSNR and SSIM are computed on the benchmark-generated outputs. Appendix A.1 describes the protocols, including evaluation on the complete WikiText-2 raw test split.
EntroPack maintains high quantization accuracy, with downstream scores such as GenEval remaining close to BF16. At 8 bpp, EntroPack achieves the lowest weight error and highest image PSNR and SSIM among the tested quantized configurations, showing that its outputs are closest to BF16. At matched bitrates near 4 bpp, EntroPack also achieves lower weight error and comparable or higher PSNR and SSIM than HRTN and HyperQuant. Its encoding and decoding are substantially faster than those of these entropy-coded alternatives.
The results across models support EntroPack’s broad applicability, while also revealing differences in quantization sensitivity. Z-Image-Turbo and Qwen3.8-27B are relatively robust, retaining good downstream performance even at 3–4 bits. The smaller Qwen3.5-9B model shows more pronounced task-performance degradation when reducing the bitrate from 4 to 3 bits.
4. Replies to “Theoretical basis of Equation 10” and “How accurate is the estimation of Equation 12?” (Q5)
Equation 10 is grounded in a theoretical bound rather than an arbitrary heuristic. We added a derivation and validation of Equation 10 in Appendix A.5. Each rANS lane starts at and ends at . The information retained across lanes per tile therefore satisfies
=\sum_{k=1}^{n_{\mathrm{tile}}}\sum_{j=1}^{J} \log_2\!\left(\frac{\xi_{k,j}}{\xi_0}\right) <16J n_{\mathrm{tile}}.$$ Since $L_{\mathrm{payload}}\approx C_{\mathrm{CE}}-I_{\mathrm{state}}$, **subtracting this upper bound gives a lower bound on the modeled payload length.** Equation 10 clips this estimate at zero and rounds it to 16-bit words. Table 8 validates it against the actual encoded payload. $$\widehat L_{\mathrm{payload}} =16\left\lceil \frac{\max\!\left(0,C_{\mathrm{CE}}-16J n_{\mathrm{tile}}\right)}{16} \right\rceil.$$ **For Equation 12, the full-matrix comparison in Table 6 shows that size estimation reduces encoding time with only a small deviation from the target bitrate.** At a 4 bpp target, it achieves 4.0178 bpp in 2.74 s, compared with 4.0000 bpp in 6.18 s using actual rANS coding. ### **5. Replies to “Tables 1 and 2 are too noisy and mixed” (Q6)** Following your suggestion, we reorganized Tables 1 and 2 into bitrate groups, making methods with similar storage budgets easier to compare. In Table 1, for example, we adjust HRTN and HyperQuant to match EntroPack’s actual bpp at its 4 bpp setting. **This organization makes the comparisons of task quality, weight reconstruction accuracy, and encoding and decoding speed clearer.**The rebuttal and the latest version of paper look good to me. The author carefully addressed my issues. I am willing to raise my score to 2. Also raise the Soundness of this version to 3. I do not consider an overall score of 3 due to the complexity of the overall method, the amount of insight and generalizable knowledge and the overall performance improvement achieved.
Thank you for taking the time to read the revised paper and our responses. Your feedback helped us strengthen the comparisons and clarify the role of each component. We’re grateful for that guidance and for your willingness to raise both scores.
Regarding your concerns about “complexity” and generalizable knowledge, EntroPack builds on a simple idea: combining quantization with entropy coding to support arbitrary target bitrates. The ablations we added in the appendix show that each design choice contributes to compression accuracy or coding efficiency. We hope the method’s “complexity” will be weighed against these demonstrated benefits.
Regarding your point about the “overall performance improvement achieved,” EntroPack delivers substantial gains in both weight reconstruction accuracy and coding speed. Compared with fixed-format methods such as INT8, FP8, and NF4, EntroPack achieves substantially lower weight reconstruction error at comparable storage. At matched bitrates, EntroPack encodes and decodes weights much faster than HRTN and HyperQuant, bringing the time cost of arbitrary-rate entropy coding much closer to that of fixed-format quantization.
We hope you will consider an overall score of 3 in light of these gains.
Thank you for identifying the gaps in the main comparisons and component ablations in our previous submission. We have added extensive evaluations and ablations addressing your concerns to better demonstrate the design principles and effectiveness of EntroPack’s pipeline. We sincerely hope that this additional evidence supports a more favorable assessment of the paper. Our responses to your main concerns are as follows.
1. Replies to “Ablations of compression-pipeline components” (Q1)
On the encoding side, we added ablation experiments in Appendix A.3, Table 6. The results at the 4 bpp target are reproduced below.
(1) Representation and coding. E8 quantization and parity reduction both provide better quantized representations; removing either substantially increases reconstruction error. Coset conditioning provides a compact coding refinement, slightly reducing both actual bpp and reconstruction error. These components have little effect on encoding time.
(2) Rate-selection cost. Subsampling and size estimation substantially accelerate encoding, at the cost of small deviations between actual and target bpp. Because they do not change the representation, they preserve the underlying rate–distortion curve and the reconstruction accuracy gained per stored bit.
(3) Post-search refinement. Row-scale refitting and RDO are both refinement steps under a fixed storage budget. The former is a simple least-squares optimization that provides a small error reduction with little effect on encoding time. RDO uses a more complex combinatorial optimization and achieves a larger error reduction, from 7.18% to 7.07%, but requires considerably more encoding time. We therefore disable RDO by default.
On the decoding side, the main factors are parallelism and kernel fusion.
(1) Parallelism. In Appendix A.8, we vary tile size to study parallelism. Smaller tiles enable more parallel decoding and improve GPU decoding speed, but add more tile metadata, requiring more storage for the same reconstruction accuracy.
(2) Kernel fusion. In Appendix A.9, we added a comparison with and without fusion of the decoding stages. Kernel fusion reduces kernel launches and global-memory reads and writes, improving decoding speed by approximately three times, as shown in Table 10.
2. Replies to “Ablations of RDO and size estimation” (Q2)
We directly ablated both components in the table above. With full-matrix search, size estimation reduces model encoding time from 6.18 s with actual rANS coding to 2.74 s. RDO instead spends additional optimization time to achieve lower weight reconstruction error at the same bitrate.
Figures 4 and 5 in Appendix A.7 further examine how sweep count and candidate scales affect encoding time and reconstruction accuracy. One sweep captures most of the improvement, while more candidate scales improve accuracy, with gains leveling off at around 20 candidates.