acceptodds
Under review as a conference paper at ICLR 2027

EntroPack: Fast and Accurate Entropy-Coded Weight Compression at Arbitrary Bitrates

Abstract

Weight compression helps large generative models fit deployment memory budgets, but common fixed-width formats offer only coarse storage choices. Entropy coding supports finer rates, yet the achieved size depends on the quantized weight distribution and coding overhead. Exploiting this flexibility requires accurate rate selection and efficient weight reconstruction for inference. We present EntroPack, an entropy-coded weight compressor that supports arbitrary target bitrates without activation calibration or fine-tuning. It combines row-normalized lattice quantization with a conditional probability model of lattice coordinates. Sampled storage estimates select the quantization resolution without repeated full-stream encoding. The final coordinates are entropy-coded in independently decodable tiles, enabling fast, fused symbol decoding and numerical weight reconstruction on the GPU. EntroPack supports floating-point and integer weight containers, including BF16, FP16, FP8, and INT8, with storage bitrate controlled independently of numerical precision. We evaluate EntroPack's compression quality and encoding speed on popular diffusion and language models. When compressing the linear-layer weights of the image generator Z-Image-Turbo, EntroPack achieves substantially lower weight and denoiser output errors than fixed-width formats at comparable storage rates, with modest denoising-step overhead. Targeting 4 bits per parameter, it achieves lower weight and denoiser output errors than NF4, including about 24% lower relative weight error, with less storage.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Weight compression, Entropy coding, Lattice quantization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1LBLOpenReview ↗
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
  1. 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.
  2. 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.
  3. 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?
  4. 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.
  5. 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?
  6. 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

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.

VariantActual bppRel. L2L_2 (%)Encode (s)
Full EntroPack (RDO off)4.02377.18341.12
No E8E_8 (scalar quantization)4.02717.74111.22
No coset conditioning (pooled tables)4.02717.19961.12
No parity reduction (direct z8z_8)4.02357.82331.34
No subsampling (full-matrix search)4.01787.21222.74
No size estimate (full-matrix rANS coding)4.00007.30066.18
No row-scale refit4.02377.20291.13
With RDO4.02377.06688.14

(1) Representation and coding. E8E_8 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.

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 ξ0=215\xi_0=2^{15} and ends at ξk,j∈[215,231)\xi_{k,j}\in[2^{15},2^{31}). The information retained across JJ 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 E8E_8 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.

Reviewer_Zx6nOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents EntroPack, a weight compression method that supports fine-grained target bitrates without activation calibration or fine-tuning. It combines row-normalized lattice quantization, coset-conditioned probability modeling, and sampled storage estimation to select quantization scales. The resulting fields are stored in independently decodable rANS tiles, with entropy decoding, lattice reconstruction, row rescaling, and dtype conversion fused on the GPU. Experiments cover image generation, audio-video generation, and language modeling. The reported results demonstrate favorable storage–weight reconstruction trade-offs relative to the tested fixed-width formats and substantially faster model conversion than the tested entropy-coded baselines.

Critical strengths
  1. The method achieves a practically useful compression trade-off without calibration data. In Table 1, EntroPack achieves a relative weight error of 7.18% at 4.02 bpp, compared with 9.41% at 4.50 bpp for NF4. Its model conversion time is 3.5 seconds, substantially below the reported 33.2 seconds for HRTN and 34.2 seconds for HyperQuant.

  2. The representation and GPU reconstruction pipeline form a coherent system design. The appendix also characterizes the relationship between tile size, decoding throughput, and storage overhead, and demonstrates compatibility with FP8/INT8 computation. These results reflect meaningful attention to practical execution.

Critical weaknesses
  1. The connection between the technical additions and the main performance gains is insufficiently established. E_8 quantization, entropy coding, tiled parallelism, and operator fusion have established precedents, so the contribution depends on demonstrating how the proposed integration produces consequential benefits. The current ablations do not sufficiently distinguish gains from sampled rate selection, the field representation, tile layout, and implementation optimizations.

  2. Task-quality and deployment evaluations are not sufficiently comprehensive. The image evaluation in Appendix A.1 uses only two prompts and three seeds, yielding six image pairs, while the video evaluation uses three seeds. PSNR and SSIM against matched BF16 outputs measure reference fidelity but do not adequately establish preservation of perceptual quality or prompt adherence.

Questions for authors
  1. Which proposed design choices account for the principal gains? Could you isolate, within the same implementation and at matched achieved bitrates, the conversion-speed benefit of sampled rate selection over full-data search, and the reconstruction-latency benefit of fusion over an unfused implementation?

  2. Does the quality comparison with HRTN and HyperQuant hold across a more diverse set of prompts at matched achieved bitrates?

  3. What is the end-to-end peak memory consumption after including uncompressed modules, decoding tables, and dense reconstruction buffers?

  4. For language modeling, how does latency change under actual draft acceptance rates or standard autoregressive decoding?

5 comments

Thank you for your constructive feedback. We have added component ablations, broader task-quality evaluations, and memory and latency analyses to address your concerns. The details are as follows.

1. Replies to “Contributions of individual technical components” (W1 & 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.

VariantActual bppRel. L2L_2 (%)Encode (s)
Full EntroPack (RDO off)4.02377.18341.12
No E8E_8 (scalar quantization)4.02717.74111.22
No coset conditioning (pooled tables)4.02717.19961.12
No parity reduction (direct z8z_8)4.02357.82331.34
No subsampling (full-matrix search)4.01787.21222.74
No size estimate (full-matrix rANS coding)4.00007.30066.18
No row-scale refit4.02377.20291.13
With RDO4.02377.06688.14

(1) Representation and coding. E8E_8 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 “Insufficient task-quality evaluations” (W2)

Following your feedback, we added extensive evaluations across multiple models, benchmarks, and metrics for image generation, video generation, and language modeling. Tables 1–5 of the revised paper report these results. Table 1, for example, shows that both 4-bit-level and 8-bit-level quantized models largely preserve image-generation quality.

The quantized models’ scores fluctuate around those of BF16, and no method achieves the highest scores across all benchmarks. Models such as Z-Image-Turbo and Qwen3.8-27B are relatively robust to quantization, retaining good task quality even at 4 bits. Here, the comparison concerns how closely each method preserves BF16’s scores, rather than which method achieves the highest score.

On GenEval and DPG-Bench, higher-rate configurations generally show smaller score variations and remain closer to BF16. PSNR and SSIM computed on images from these benchmarks further show that EntroPack at 8 bpp produces outputs closest to BF16, supporting its higher compression fidelity.

3. Replies to “Comparison with HRTN and HyperQuant at matched achieved bitrates” (Q2)

We adopted your suggested comparison in the main quality evaluations, adjusting the target bpp of HRTN and HyperQuant so that their actual bpp matches EntroPack’s. At matched bitrates, EntroPack consistently achieves lower weight error and comparable or higher PSNR and SSIM. Its encoding and decoding are also substantially faster than those of these two methods.

4. Replies to “End-to-end peak memory consumption” (Q3)

We added the peak GPU-memory usage of every method in Tables 1–4, including the uncompressed modules, decoding tables, and dense reconstruction buffers you mentioned. These measurements directly reflect memory requirements during deployment and generally show that smaller compressed weights lead to lower peak memory. In Table 1, EntroPack at 3 bpp achieves the lowest peak memory, using less than one third of the BF16 baseline.

5. Replies to “Language-model latency across workloads” (Q4)

We added detailed latency analyses in Appendix A.10 for LLM prefill, batch decoding, and multi-token prediction.

Across these workloads, each forward pass comprises weight decoding and model computation. EntroPack’s weight decoding for Qwen3.8-27B at 4 bpp consistently takes about 104 ms, as reflected throughout Figure 7. Longer prefills, larger batches, and multi-token prediction allow more tokens to share this decoding cost, improving overall inference efficiency.

Dear Reviewer Zx6n,

Thank you again for your valuable comments. As the discussion period closes in two days, we would like to kindly check whether our response and revision have addressed your concerns. We would be grateful for any feedback and are happy to clarify anything further.

Best regards, The Authors

Reviewer_KE6UOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces EntroPack, an entropy-coded weight compression method that supports fine-grained target bitrates without activation calibration or fine-tuning. It combines row-normalized E8 lattice quantization, conditional entropy modeling, and sampled storage estimation for rate selection, with tiled coding for fused GPU decoding and weight reconstruction. Experiments on diffusion and language models report competitive reconstruction quality and fast compression, with workload-dependent inference overhead.

Critical strengths
  1. Clear motivation and presentation. The paper clearly motivates the need for fine-grained weight storage and identifies two practical challenges: selecting a quantization resolution that meets the target bitrate and efficiently reconstructing weights for inference. The proposed storage estimation and fused GPU reconstruction directly address these challenges.

  2. Broad coverage of evaluation dimensions. The experiments assess actual storage, weight reconstruction error, model output fidelity, compression time, and inference latency across diffusion and language models. This provides a useful view of the tradeoffs between compression quality and computational cost.

Critical weaknesses
  1. Limited evaluation of downstream model quality. The image fidelity results in Table 1 use only two prompts and three random seeds. Moreover, PSNR and SSIM measure similarity to the uncompressed model’s outputs rather than overall generation quality or prompt adherence. The language-model evaluation primarily reports WikiText-2 perplexity, leaving performance on broader downstream tasks unclear. Larger-scale generation evaluations and additional language tasks would strengthen the evidence that compression preserves practical model capabilities.

  2. Substantial inference overhead in the reported language-model decoding workload. Although the overhead is modest for diffusion denoising and language-model prefill, Table 2 reports an increase from 155.1 ms to 272.8 ms for the specified eight-token decoding workload, approximately 76% above BF16. This limits the practical benefit in latency-sensitive generation settings. A more detailed evaluation across decoding workloads, together with peak memory measurements, would help clarify when the storage savings justify the reconstruction overhead.

Questions for authors

Table 2 reports a latency increase from 155.1 ms to 272.8 ms for the specified eight-token decoding workload. How does this overhead change with batch size, context length, and standard single-token autoregressive decoding? Could you separate weight reconstruction time from the remaining inference time?

see weakness.

4 comments

Thank you for your constructive feedback. We have expanded the downstream evaluations and added inference-time and memory analyses to address your concerns.

1. Replies to “Limited evaluation of downstream model quality” (W1)

We have expanded the evaluation beyond reference-output similarity. For image generation, we evaluate two models on the widely used GenEval and DPG-Bench benchmarks, with 6,472 generated images per row in Tables 1 and 3, and report PickScore, PSNR, and SSIM in addition to the benchmark scores. For three LLMs, we supplement WikiText-2 perplexity with MMLU, GSM8K, HellaSwag, and ARC-Challenge in Tables 2 and 4.

These results show that EntroPack largely preserves downstream task performance across the evaluated models. Appendix A.1 of the revised paper provides the evaluation protocols.

2. Replies to “Inference overhead in language-model decoding” (W2 & Q1)

We added inference-time breakdowns in Appendix A.10 to analyze reconstruction overhead in single-token decoding, batching, multi-token prediction, and prefilling with varying context lengths. These results indicate that EntroPack offers greater advantages in compute-intensive workloads, including diffusion denoising, LLM prefill, and LLM training.

Under a constrained GPU-memory budget, EntroPack can translate memory savings into improved weight accuracy and faster inference through quantization and flexible bitrate settings, as shown in Table 12.

We acknowledge that weight reconstruction remains a substantial overhead for single-token autoregressive decoding, but this overhead can be amortized through multi-token prediction and batch prediction, as shown in Figure 7.

Dear Reviewer KE6U,

Thank you again for your valuable comments. As the discussion period closes in two days, we would like to kindly check whether our response and revision have addressed your concerns. We would be grateful for any feedback and are happy to clarify anything further.

Best regards, The Authors

Thank you for the detailed rebuttal and clarifications. I will maintain my original rating.

Dear Reviewer KE6U,

Thank you very much for your positive feedback. We sincerely appreciate your time and constructive comments, which have helped improve our paper. We will incorporate the corresponding revisions into the final version.

Best regards, The Authors

Comments on the paper

1 comment

Dear Reviewers and Area Chairs,

Thank you for your constructive feedback. Following the reviewers’ comments, we have substantially expanded the experiments and theoretical explanations, and further optimized EntroPack’s encoding speed. These additions strengthen the evidence for its quantization accuracy, quality preservation, and encoding/decoding efficiency. We sincerely hope that these additions address your concerns and that reviewers will reconsider their ratings in light of the new evidence and revised paper.

The main revisions are as follows:

  1. Extensive model and benchmark evaluation. Our expanded experiments demonstrate EntroPack’s favorable weight reconstruction accuracy and preservation of task quality, together with substantially faster encoding and decoding than the tested entropy-coded baselines. These experiments cover Z-Image-Turbo, Qwen-Image-2.1-Turbo, MiniMax-H3, and three LLMs: Qwen3.8-27B, Muse-Glimmer-30B, and Qwen3.5-9B. We evaluate these models on widely used benchmarks in their respective fields, including GenEval, DPG-Bench, VBench, and five language benchmarks.

  2. Comprehensive ablations and analysis. We added systematic component ablations to isolate the contribution of each design choice in representation and coding, rate selection, and post-search refinement. Additional analyses cover RDO, size estimation and its theoretical basis, row subsampling, tile-level parallelism, kernel fusion, and inference time breakdowns.

  3. New encoding optimizations. We further optimized EntroPack’s size-estimation implementation, achieving several-fold speedups in total model encoding without changing the compressed outputs. We report these new results in Appendix A.4 and have uploaded both the pre-optimization and optimized implementations as supplementary code.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.